This notebook was created by Jean de Dieu Nyandwi for the love of machine learning community. For any feedback, errors or suggestion, he can be reached on email (johnjw7084 at gmail dot com), Twitter, or LinkedIn.

CNN for Real World Dataset and Image Augmentation¶

Contents¶

• 1. Intro - Real World Datasets and Data Augmentation

• 2. Getting Started: Real World Datasets and Overfitting

  • 2.1 Loading and Preparing the Data
  • 2.2 Buiding, Compiling and Training a Model
  • 2.3 Visualizing the Model Results

• 3. Image Augmentation with ImageDataGenerator

  • 3.1 Loading the Data Again
  • 3.2 Applying Data Augmentation
  • 3.3 Visualizing Augmented Images
  • 3.4 Retraining a Model on Augmented Data
  • 3.5 Visualizing the Model Resuts
  • 3.6 Further Improvements: Using Pretrained Models
  • 3.7 Saving, Loading, and Testing a Model on New Images

• 4. Image Augmentation with Keras Image Augmentation Layers

  • 4.1 Getting the Data again
  • 4.2 Looking in the Data
  • 4.3 Preprocessing Data
  • 4.4 Augmenting the Images with Auggmentation Layers
  • 4.5 First Option: Building a Model with an Augmented Images
  • 4.6 Second Option: Augmenting the Images Inside a Model

• 5. Further Learning

1. Intro - Real World Datasets and Image Augmentation¶

Real world image datasets are not always prepared and most of the time, they are not enough in quantity.

Training an effective computer vision system requires a huge amount of images. That is not always the case though. When MIT Technology Review asked Andrew Ng. about the size of the data required to build an AI project, he said:

"Machine learning is so diverse that it’s become really hard to give one-size-fits-all answers. I’ve worked on problems where I had about 200 to 300 million images. I’ve also worked on problems where I had 10 images, and everything in between. When I look at manufacturing applications, I think something like tens or maybe a hundred images for a defect class is not unusual, but there’s very wide variance even within the factory".

We can not emphasize enough the advantages of having both good and enough data although such blend is not always possible.

Given that sometime we may have a handful of images, how can we go about it? Are there ways to expand the small image dataset? And boost the performance of the machine learning model as a result?

With the advancement of machine learning techniques, it has become possible to add a boost to the performance metrics by just expanding the existing dataset. The technique of synthetizing new data(images) from existing data is called data augmentation.

image datasets can be augmented in various ways including:

• Flipping the image, vertically or horizontally.
• Cropping the image.
• Changing contrast and color of the image.
• Adding noise to the data.
• Rotating the image at a given degree.

Below image summarize all possibilities that can be done in image data augmentation.

2. Getting Started: Real World Datasets and Overfitting¶

One of the most challenges in training machine learning models on real world datasets is overfitting.

A model overfits when it memorized the training data due to insufficient training samples, or lack of diversity in training samples. By doing data augmentation, we are increasing the training samples, as well as introducing some diversity in the images.

It's fair to say that data augmentation is the cure to overfitting. To test that, let's train a quick cat and dog classifier. After that, we will augment the images to ovoid overfitting.

Without Data Augmentation: Training Cat and Dog Classifier¶

We are going to see the need of data augmentation by training a cat and dog classifier.

Along the way, we will let the results guide the latter.

2.1 Loading and Preparing Cat and Dog Data¶

Imports¶

In [ ]:

import tensorflow as tf
from tensorflow import keras
import os 
import zipfile
import matplotlib.pyplot as plt
import numpy as np

import tensorflow as tf from tensorflow import keras import os import zipfile import matplotlib.pyplot as plt import numpy as np

Getting the data¶

The version of the data that we are going to use here is a filtered version. Orginally, it contains over 20.000 images.

In [ ]:

# Load the data into the workspace

!wget --no-check-certificate \
    https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip \
    -O /tmp/cats_and_dogs_filtered.zip

# Load the data into the workspace !wget --no-check-certificate \ https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip \ -O /tmp/cats_and_dogs_filtered.zip

--2021-09-18 06:21:48--  https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip
Resolving storage.googleapis.com (storage.googleapis.com)... 209.85.147.128, 142.250.125.128, 142.250.136.128, ...
Connecting to storage.googleapis.com (storage.googleapis.com)|209.85.147.128|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 68606236 (65M) [application/zip]
Saving to: ‘/tmp/cats_and_dogs_filtered.zip’

/tmp/cats_and_dogs_ 100%[===================>]  65.43M   178MB/s    in 0.4s    

2021-09-18 06:21:48 (178 MB/s) - ‘/tmp/cats_and_dogs_filtered.zip’ saved [68606236/68606236]

In [ ]:

# Extract the zip file

zip_dir = '/tmp/cats_and_dogs_filtered.zip'
zip_ref = zipfile.ZipFile(zip_dir, 'r')
zip_ref.extractall('/tmp')
zip_ref.close()

# Extract the zip file zip_dir = '/tmp/cats_and_dogs_filtered.zip' zip_ref = zipfile.ZipFile(zip_dir, 'r') zip_ref.extractall('/tmp') zip_ref.close()

Now that the images are extracted (you can verify that by looking in Colab directory), let's get the training and validation directories from the main directory.

In [ ]:

main_dir = '/tmp/cats_and_dogs_filtered'

train_dir = os.path.join(main_dir, 'train')
val_dir = os.path.join(main_dir, 'validation')

main_dir = '/tmp/cats_and_dogs_filtered' train_dir = os.path.join(main_dir, 'train') val_dir = os.path.join(main_dir, 'validation')

We will continue with getting the cats and dogs in the above defined directories.

In [ ]:

train_cat_dir = os.path.join(train_dir, 'cats')
train_dog_dir = os.path.join(train_dir, 'dogs')

val_cat_dir = os.path.join(val_dir, 'cats')
val_dog_dir = os.path.join(val_dir, 'dogs')

train_cat_dir = os.path.join(train_dir, 'cats') train_dog_dir = os.path.join(train_dir, 'dogs') val_cat_dir = os.path.join(val_dir, 'cats') val_dog_dir = os.path.join(val_dir, 'dogs')

Now, our directories are quite arranged. Let's see peep into the directories.

In [ ]:

os.listdir(main_dir)

os.listdir(main_dir)

Out[ ]:

['vectorize.py', 'train', 'validation']

In [ ]:

os.listdir(train_cat_dir)[:10]

os.listdir(train_cat_dir)[:10]

Out[ ]:

['cat.504.jpg',
 'cat.148.jpg',
 'cat.560.jpg',
 'cat.506.jpg',
 'cat.182.jpg',
 'cat.300.jpg',
 'cat.178.jpg',
 'cat.791.jpg',
 'cat.260.jpg',
 'cat.328.jpg']

Let's walk through the directories and see the number of images in each directory.

In [ ]:

for dir, dirname, filename in os.walk(main_dir):
  print(f"Found {len(dirname)} directories and {len(filename)} images in {dir}")

for dir, dirname, filename in os.walk(main_dir): print(f"Found {len(dirname)} directories and {len(filename)} images in {dir}")

Found 2 directories and 1 images in /tmp/cats_and_dogs_filtered
Found 2 directories and 0 images in /tmp/cats_and_dogs_filtered/train
Found 0 directories and 1000 images in /tmp/cats_and_dogs_filtered/train/dogs
Found 0 directories and 1000 images in /tmp/cats_and_dogs_filtered/train/cats
Found 2 directories and 0 images in /tmp/cats_and_dogs_filtered/validation
Found 0 directories and 500 images in /tmp/cats_and_dogs_filtered/validation/dogs
Found 0 directories and 500 images in /tmp/cats_and_dogs_filtered/validation/cats

There are 3000 images, 2000 in training set, 1000 in validation set. Cats and dogs images are evenly splitted.

Let's now generate a dataset to train the model.

Preparing the Dataset¶

In [ ]:

from keras.preprocessing.image import ImageDataGenerator

# Rescale the image to values between 0 and 1

train_gen = ImageDataGenerator(rescale=1/255.0)
val_gen = ImageDataGenerator(rescale=1/255.0)

batch_size = 20
image_size = (180,180)

train_data = train_gen.flow_from_directory(train_dir, 
                                           batch_size=batch_size,
                                           class_mode='binary',
                                           target_size=image_size)

val_data = val_gen.flow_from_directory(val_dir, 
                                           batch_size=batch_size,
                                           class_mode='binary',
                                           target_size=image_size)

from keras.preprocessing.image import ImageDataGenerator # Rescale the image to values between 0 and 1 train_gen = ImageDataGenerator(rescale=1/255.0) val_gen = ImageDataGenerator(rescale=1/255.0) batch_size = 20 image_size = (180,180) train_data = train_gen.flow_from_directory(train_dir, batch_size=batch_size, class_mode='binary', target_size=image_size) val_data = val_gen.flow_from_directory(val_dir, batch_size=batch_size, class_mode='binary', target_size=image_size)

Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.

Before building a model, let's visualize the images. It's always a best practice.

In [ ]:

data_for_viz = tf.keras.preprocessing.image_dataset_from_directory(
    
          train_dir,
          image_size=(180,180))

data_for_viz = tf.keras.preprocessing.image_dataset_from_directory( train_dir, image_size=(180,180))

Found 2000 files belonging to 2 classes.

In [ ]:

def image_viz(dataset):
  
  plt.figure(figsize=(12, 8))
  
  index = 0
  for image, label in dataset.take(12):
    
    index +=1
    ax = plt.subplot(4, 4, index)
    plt.imshow(image[index].numpy().astype("uint8"))
    plt.title(int(label[index]))
    plt.axis("off")

def image_viz(dataset): plt.figure(figsize=(12, 8)) index = 0 for image, label in dataset.take(12): index +=1 ax = plt.subplot(4, 4, index) plt.imshow(image[index].numpy().astype("uint8")) plt.title(int(label[index])) plt.axis("off")

In [ ]:

image_viz(data_for_viz)

image_viz(data_for_viz)

2.2 Building, Compiling and Training a Model¶

In [ ]:

def classifier():

  model = tf.keras.models.Sequential([

    # First convolution and pooling layer                                  
    tf.keras.layers.Conv2D(32, (3,3), activation='relu', input_shape=(180, 180, 3)),
    tf.keras.layers.MaxPooling2D(2, 2),
    # Second convolution and pooling layer

    tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
    tf.keras.layers.MaxPooling2D(2,2),

    # Third convolution and pooling layer

    tf.keras.layers.Conv2D(128, (3,3), activation='relu'),
    tf.keras.layers.MaxPooling2D(2,2),

    # Flattening layer for converting the feature maps into 1D column vector
    
    tf.keras.layers.Flatten(),

    # Fully connected layers

    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(1, activation='sigmoid')

  ])


  # Compiling the model: Specifying optimizer, loss and metric to track during training

  model.compile(
      
      optimizer=tf.keras.optimizers.RMSprop(),
      loss='binary_crossentropy',
      metrics=['accuracy']
  )

  return model

def classifier(): model = tf.keras.models.Sequential([ # First convolution and pooling layer tf.keras.layers.Conv2D(32, (3,3), activation='relu', input_shape=(180, 180, 3)), tf.keras.layers.MaxPooling2D(2, 2), # Second convolution and pooling layer tf.keras.layers.Conv2D(64, (3,3), activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Third convolution and pooling layer tf.keras.layers.Conv2D(128, (3,3), activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Flattening layer for converting the feature maps into 1D column vector tf.keras.layers.Flatten(), # Fully connected layers tf.keras.layers.Dense(128, activation='relu'), tf.keras.layers.Dense(1, activation='sigmoid') ]) # Compiling the model: Specifying optimizer, loss and metric to track during training model.compile( optimizer=tf.keras.optimizers.RMSprop(), loss='binary_crossentropy', metrics=['accuracy'] ) return model

In [ ]:

# Training a model

model = classifier()

train_steps = 2000 / batch_size
val_steps = 1000 / batch_size

history = model.fit(train_data, 
          validation_data=val_data,
          epochs=50,
          steps_per_epoch=train_steps,
          validation_steps=val_steps)

# Training a model model = classifier() train_steps = 2000 / batch_size val_steps = 1000 / batch_size history = model.fit(train_data, validation_data=val_data, epochs=50, steps_per_epoch=train_steps, validation_steps=val_steps)

Epoch 1/50
100/100 [==============================] - 45s 121ms/step - loss: 0.8243 - accuracy: 0.5435 - val_loss: 0.7136 - val_accuracy: 0.5000
Epoch 2/50
100/100 [==============================] - 12s 118ms/step - loss: 0.6378 - accuracy: 0.6455 - val_loss: 0.5870 - val_accuracy: 0.6690
Epoch 3/50
100/100 [==============================] - 12s 118ms/step - loss: 0.5446 - accuracy: 0.7245 - val_loss: 0.5848 - val_accuracy: 0.6990
Epoch 4/50
100/100 [==============================] - 12s 117ms/step - loss: 0.4387 - accuracy: 0.7955 - val_loss: 0.5780 - val_accuracy: 0.7310
Epoch 5/50
100/100 [==============================] - 12s 117ms/step - loss: 0.3505 - accuracy: 0.8525 - val_loss: 0.5507 - val_accuracy: 0.7550
Epoch 6/50
100/100 [==============================] - 12s 119ms/step - loss: 0.2380 - accuracy: 0.8940 - val_loss: 0.7112 - val_accuracy: 0.7190
Epoch 7/50
100/100 [==============================] - 12s 118ms/step - loss: 0.1427 - accuracy: 0.9425 - val_loss: 0.8400 - val_accuracy: 0.7560
Epoch 8/50
100/100 [==============================] - 12s 118ms/step - loss: 0.1045 - accuracy: 0.9620 - val_loss: 0.9308 - val_accuracy: 0.7460
Epoch 9/50
100/100 [==============================] - 12s 116ms/step - loss: 0.0593 - accuracy: 0.9785 - val_loss: 1.2529 - val_accuracy: 0.7140
Epoch 10/50
100/100 [==============================] - 12s 119ms/step - loss: 0.0575 - accuracy: 0.9815 - val_loss: 1.5428 - val_accuracy: 0.7390
Epoch 11/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0421 - accuracy: 0.9870 - val_loss: 1.9313 - val_accuracy: 0.7350
Epoch 12/50
100/100 [==============================] - 12s 117ms/step - loss: 0.0349 - accuracy: 0.9875 - val_loss: 1.9470 - val_accuracy: 0.7300
Epoch 13/50
100/100 [==============================] - 12s 117ms/step - loss: 0.0599 - accuracy: 0.9875 - val_loss: 2.0286 - val_accuracy: 0.7350
Epoch 14/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0267 - accuracy: 0.9925 - val_loss: 2.0821 - val_accuracy: 0.7330
Epoch 15/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0440 - accuracy: 0.9920 - val_loss: 2.3201 - val_accuracy: 0.7280
Epoch 16/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0649 - accuracy: 0.9885 - val_loss: 2.2570 - val_accuracy: 0.7180
Epoch 17/50
100/100 [==============================] - 12s 121ms/step - loss: 0.0521 - accuracy: 0.9930 - val_loss: 2.5457 - val_accuracy: 0.7230
Epoch 18/50
100/100 [==============================] - 12s 119ms/step - loss: 0.0297 - accuracy: 0.9950 - val_loss: 2.4022 - val_accuracy: 0.7310
Epoch 19/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0482 - accuracy: 0.9900 - val_loss: 2.2503 - val_accuracy: 0.7210
Epoch 20/50
100/100 [==============================] - 12s 120ms/step - loss: 0.0227 - accuracy: 0.9945 - val_loss: 2.8755 - val_accuracy: 0.7150
Epoch 21/50
100/100 [==============================] - 12s 117ms/step - loss: 0.0265 - accuracy: 0.9950 - val_loss: 2.5832 - val_accuracy: 0.7010
Epoch 22/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0241 - accuracy: 0.9930 - val_loss: 2.5352 - val_accuracy: 0.7140
Epoch 23/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0132 - accuracy: 0.9980 - val_loss: 3.2614 - val_accuracy: 0.7200
Epoch 24/50
100/100 [==============================] - 12s 117ms/step - loss: 0.0311 - accuracy: 0.9945 - val_loss: 3.4943 - val_accuracy: 0.7230
Epoch 25/50
100/100 [==============================] - 12s 116ms/step - loss: 0.0894 - accuracy: 0.9910 - val_loss: 3.6839 - val_accuracy: 0.7070
Epoch 26/50
100/100 [==============================] - 12s 115ms/step - loss: 0.0231 - accuracy: 0.9940 - val_loss: 3.8832 - val_accuracy: 0.6890
Epoch 27/50
100/100 [==============================] - 12s 117ms/step - loss: 1.7756e-04 - accuracy: 1.0000 - val_loss: 4.3875 - val_accuracy: 0.7130
Epoch 28/50
100/100 [==============================] - 11s 115ms/step - loss: 0.0113 - accuracy: 0.9960 - val_loss: 4.3880 - val_accuracy: 0.7220
Epoch 29/50
100/100 [==============================] - 12s 117ms/step - loss: 0.0622 - accuracy: 0.9955 - val_loss: 3.6404 - val_accuracy: 0.7020
Epoch 30/50
100/100 [==============================] - 12s 117ms/step - loss: 0.0082 - accuracy: 0.9970 - val_loss: 6.3984 - val_accuracy: 0.6690
Epoch 31/50
100/100 [==============================] - 12s 116ms/step - loss: 0.0135 - accuracy: 0.9965 - val_loss: 4.5262 - val_accuracy: 0.7060
Epoch 32/50
100/100 [==============================] - 12s 117ms/step - loss: 0.0279 - accuracy: 0.9970 - val_loss: 4.9610 - val_accuracy: 0.7290
Epoch 33/50
100/100 [==============================] - 14s 136ms/step - loss: 0.0084 - accuracy: 0.9975 - val_loss: 4.2445 - val_accuracy: 0.7160
Epoch 34/50
100/100 [==============================] - 12s 120ms/step - loss: 0.0608 - accuracy: 0.9915 - val_loss: 5.2009 - val_accuracy: 0.7000
Epoch 35/50
100/100 [==============================] - 12s 119ms/step - loss: 0.0191 - accuracy: 0.9960 - val_loss: 4.1152 - val_accuracy: 0.7220
Epoch 36/50
100/100 [==============================] - 12s 117ms/step - loss: 0.0161 - accuracy: 0.9970 - val_loss: 5.3480 - val_accuracy: 0.7060
Epoch 37/50
100/100 [==============================] - 12s 119ms/step - loss: 0.0276 - accuracy: 0.9950 - val_loss: 4.9586 - val_accuracy: 0.7090
Epoch 38/50
100/100 [==============================] - 12s 119ms/step - loss: 0.0266 - accuracy: 0.9960 - val_loss: 4.8451 - val_accuracy: 0.7210
Epoch 39/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0117 - accuracy: 0.9985 - val_loss: 4.7765 - val_accuracy: 0.7290
Epoch 40/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0169 - accuracy: 0.9985 - val_loss: 5.0485 - val_accuracy: 0.7170
Epoch 41/50
100/100 [==============================] - 12s 118ms/step - loss: 3.9937e-07 - accuracy: 1.0000 - val_loss: 5.2801 - val_accuracy: 0.7220
Epoch 42/50
100/100 [==============================] - 12s 119ms/step - loss: 1.4917e-08 - accuracy: 1.0000 - val_loss: 5.5378 - val_accuracy: 0.7260
Epoch 43/50
100/100 [==============================] - 12s 119ms/step - loss: 1.6792e-08 - accuracy: 1.0000 - val_loss: 6.1406 - val_accuracy: 0.7250
Epoch 44/50
100/100 [==============================] - 12s 117ms/step - loss: 0.1762 - accuracy: 0.9930 - val_loss: 8.8898 - val_accuracy: 0.7230
Epoch 45/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0034 - accuracy: 0.9995 - val_loss: 6.9045 - val_accuracy: 0.7220
Epoch 46/50
100/100 [==============================] - 12s 119ms/step - loss: 0.0520 - accuracy: 0.9960 - val_loss: 6.7058 - val_accuracy: 0.7110
Epoch 47/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0240 - accuracy: 0.9960 - val_loss: 6.5657 - val_accuracy: 0.7270
Epoch 48/50
100/100 [==============================] - 12s 119ms/step - loss: 0.0397 - accuracy: 0.9950 - val_loss: 5.1782 - val_accuracy: 0.7260
Epoch 49/50
100/100 [==============================] - 12s 118ms/step - loss: 2.9981e-07 - accuracy: 1.0000 - val_loss: 6.1723 - val_accuracy: 0.7270
Epoch 50/50
100/100 [==============================] - 12s 118ms/step - loss: 0.0144 - accuracy: 0.9975 - val_loss: 5.6709 - val_accuracy: 0.7110

2.3 Visualizing the Model Results¶

In [17]:

# function to plot accuracy and loss

def plot_acc_loss(acc, val_acc, loss, val_loss, epochs):

  plt.figure(figsize=(10,5))
  plt.plot(epochs, acc, 'r', label='Training Accuracy')
  plt.plot(epochs, val_acc, 'g', label='Validation Accuracy')
  plt.title('Training and validation accuracy')
  plt.xlabel('Epochs')
  plt.ylabel('Accuracy')
  plt.legend(loc=0)

  # Create a new figure with plt.figure()
  plt.figure()

  plt.figure(figsize=(10,5))
  plt.plot(epochs, loss, 'b', label='Training Loss')
  plt.plot(epochs, val_loss, 'y', label='Validation Loss')
  plt.title('Training and Validation Loss')
  plt.xlabel('Epochs')
  plt.ylabel('Loss')
  plt.legend(loc=0)
  plt.show()

# function to plot accuracy and loss def plot_acc_loss(acc, val_acc, loss, val_loss, epochs): plt.figure(figsize=(10,5)) plt.plot(epochs, acc, 'r', label='Training Accuracy') plt.plot(epochs, val_acc, 'g', label='Validation Accuracy') plt.title('Training and validation accuracy') plt.xlabel('Epochs') plt.ylabel('Accuracy') plt.legend(loc=0) # Create a new figure with plt.figure() plt.figure() plt.figure(figsize=(10,5)) plt.plot(epochs, loss, 'b', label='Training Loss') plt.plot(epochs, val_loss, 'y', label='Validation Loss') plt.title('Training and Validation Loss') plt.xlabel('Epochs') plt.ylabel('Loss') plt.legend(loc=0) plt.show()

In [ ]:

model_history = history.history

acc = model_history['accuracy']
val_acc = model_history['val_accuracy']
loss = model_history['loss']
val_loss = model_history['val_loss']

epochs = history.epoch

model_history = history.history acc = model_history['accuracy'] val_acc = model_history['val_accuracy'] loss = model_history['loss'] val_loss = model_history['val_loss'] epochs = history.epoch

In [ ]:

plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

<Figure size 432x288 with 0 Axes>

Clearly, the our classifier overfitted. If you can look at the plots above, there is so much gap between the training and validation accuracy/loss.

The classifier is overly confident at recognizing training images, but not so good when evaluated on the validation images.

This goes to show that data augmentation is very useful technique. By just expanding the images, and introducing different image scenes, overfitting can potentially be handled. That is what we are going to do iin the next section.

3. Image Augmentation with ImageDataGenerator¶

ImageDataGenerator is a powerful Keras image processing functionality used to augment images. It is a part of image data processing functions.

The single most advantage of ImageDataGenerator is that it allows you to augment images in realtime, as you load them from a directory for example.

The orginal directory of the data is not affected at all. The image will be loaded & augmented at the same time, while not affecting the orginal directory.

3.1 Loading the Data Again¶

In [ ]:

# Download the data into the workspace

!wget --no-check-certificate \
    https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip \
    -O /tmp/cats_and_dogs_filtered.zip


# Extract the zip file

zip_dir = '/tmp/cats_and_dogs_filtered.zip'
zip_ref = zipfile.ZipFile(zip_dir, 'r')
zip_ref.extractall('/tmp')
zip_ref.close()

# Get training and val directories

main_dir = '/tmp/cats_and_dogs_filtered'

train_dir = os.path.join(main_dir, 'train')
val_dir = os.path.join(main_dir, 'validation')

train_cat_dir = os.path.join(train_dir, 'cats')
train_dog_dir = os.path.join(train_dir, 'dogs')

val_cat_dir = os.path.join(val_dir, 'cats')
val_dog_dir = os.path.join(val_dir, 'dogs')

# Download the data into the workspace !wget --no-check-certificate \ https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip \ -O /tmp/cats_and_dogs_filtered.zip # Extract the zip file zip_dir = '/tmp/cats_and_dogs_filtered.zip' zip_ref = zipfile.ZipFile(zip_dir, 'r') zip_ref.extractall('/tmp') zip_ref.close() # Get training and val directories main_dir = '/tmp/cats_and_dogs_filtered' train_dir = os.path.join(main_dir, 'train') val_dir = os.path.join(main_dir, 'validation') train_cat_dir = os.path.join(train_dir, 'cats') train_dog_dir = os.path.join(train_dir, 'dogs') val_cat_dir = os.path.join(val_dir, 'cats') val_dog_dir = os.path.join(val_dir, 'dogs')

--2021-09-18 06:32:33--  https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip
Resolving storage.googleapis.com (storage.googleapis.com)... 74.125.70.128, 74.125.201.128, 74.125.202.128, ...
Connecting to storage.googleapis.com (storage.googleapis.com)|74.125.70.128|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 68606236 (65M) [application/zip]
Saving to: ‘/tmp/cats_and_dogs_filtered.zip’

/tmp/cats_and_dogs_ 100%[===================>]  65.43M   169MB/s    in 0.4s    

2021-09-18 06:32:34 (169 MB/s) - ‘/tmp/cats_and_dogs_filtered.zip’ saved [68606236/68606236]

3.2 Apply Data Augmentation¶

We are going to use ImageDataGenerator to generate augmented images.

Below are some of the options available in ImageDataGenerator and their explainations.

train_imagenerator = ImageDataGenerator(
      rotation_range=40,
      width_shift_range=0.2,
      height_shift_range=0.2,
      shear_range=0.2,
      zoom_range=0.2,
      horizontal_flip=True,
      fill_mode='nearest')

• rotation_range is a value in degrees (0–180) to randomly rotate images.

• width_shift and height_shift are ranges of fraction of total width or height within which to translate pictures, either vertically or horizontally.

• shear_range is for applying shearing randomly.

• zoom_range is for zooming pictures randomly.

• horizontal_flip is for flipping half of the images horizontally. There is also vertical_flip option.

• fill_mode is for completing newly created pixels, which can appear after a rotation or a width/height shift.

See the documentation, it is an interesting read, and there are more preprocessing functions that you might need in your future projects.

Let's see this in practice! We will start by creating train_imagenerator which is an image generator for training set.

In [ ]:

# Creating training image data generator

from tensorflow.keras.preprocessing.image import ImageDataGenerator

train_imagenerator = ImageDataGenerator(
    
    rescale=1/255.,
    rotation_range=40,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
    fill_mode='nearest'
    
)

# Creating training image data generator from tensorflow.keras.preprocessing.image import ImageDataGenerator train_imagenerator = ImageDataGenerator( rescale=1/255., rotation_range=40, width_shift_range=0.2, height_shift_range=0.2, shear_range=0.2, zoom_range=0.2, horizontal_flip=True, fill_mode='nearest' )

We will also create val_imagenerator, but different to training generator, there are no augmentation. It's only rescaling the images pixels to values between 0 and 1. Images pixels are normally between 0 and 255. Rescaling the values improve the performance of the neural network and reduce training time as well.

In [ ]:

# Validation image generator

val_imagenerator = ImageDataGenerator(rescale=1/255.)

# Validation image generator val_imagenerator = ImageDataGenerator(rescale=1/255.)

After creating train and validation generators, let's use image_dataset_from_directory function to generate a TensorFlow dataset from images files located in our two directories.

Below is how training and validation directories are structured:

cats_dogs_filtered/
..train/
....cats/
......cat.0.jpg
......cat.1.jpg
......cat.2.jpg
....dogs/
......dog.0.jpg
......dog.1.jpg
......dog.2.jpg
..validation/
....cats/
......cat.0.jpg
......cat.1.jpg
....dogs/
......dog.0.jpg
......dog.1.jpg

In [ ]:

# Load training images in batches of 20 while applying aumgmentation

batch_size = 20
target_size = (180,180)

train_generator = train_imagenerator.flow_from_directory(
        train_dir, #parent directory must be specified
        target_size = target_size, # All images will be resized to (180,180)
        batch_size=batch_size,
        class_mode='binary' # since we need binary labels(0,1) and we will use binary_crossentropy

)

# Load training images in batches of 20 while applying aumgmentation batch_size = 20 target_size = (180,180) train_generator = train_imagenerator.flow_from_directory( train_dir, #parent directory must be specified target_size = target_size, # All images will be resized to (180,180) batch_size=batch_size, class_mode='binary' # since we need binary labels(0,1) and we will use binary_crossentropy )

Found 2000 images belonging to 2 classes.

In [ ]:

val_generator = val_imagenerator.flow_from_directory(
        val_dir, #parent directory must be specified
        target_size = target_size, # All images will be resized to (180,180)
        batch_size=batch_size,
        class_mode='binary' # since we need binary labels(0,1) and we will use binary_crossentropy

)

val_generator = val_imagenerator.flow_from_directory( val_dir, #parent directory must be specified target_size = target_size, # All images will be resized to (180,180) batch_size=batch_size, class_mode='binary' # since we need binary labels(0,1) and we will use binary_crossentropy )

Found 1000 images belonging to 2 classes.

Another great advantage of ImageDataGenerator is that it generates the labels of the images based off their folders. During the model training, we won't need to specify the labels.

We are ready to train our machine learning model now, but before we could try to visualize the augmented images.

3.3 Visualizing Augmented Images¶

In [ ]:

# Get images in batch of 20
augmented_image, label = train_generator.next()

plt.figure(figsize=(12,8))
for i in range(6):
        ax = plt.subplot(2, 3, i + 1)
        plt.imshow(augmented_image[i])
        plt.title(int(label[i]))
        plt.axis("off")

# Get images in batch of 20 augmented_image, label = train_generator.next() plt.figure(figsize=(12,8)) for i in range(6): ax = plt.subplot(2, 3, i + 1) plt.imshow(augmented_image[i]) plt.title(int(label[i])) plt.axis("off")

It may be hard to spot since we are not comparing them with non augmented images, but if you can observe well, some images are zoomed in, rotated, and flipped horizantally.

Below images are not augmented. You can see that no zoom were applied for example.

In [ ]:

non_augmented_image, label = train_data.next()

plt.figure(figsize=(12,8))
for i in range(6):
        ax = plt.subplot(2, 3, i + 1)
        plt.imshow(non_augmented_image[i])
        plt.title(int(label[i]))
        plt.axis("off")

non_augmented_image, label = train_data.next() plt.figure(figsize=(12,8)) for i in range(6): ax = plt.subplot(2, 3, i + 1) plt.imshow(non_augmented_image[i]) plt.title(int(label[i])) plt.axis("off")

Now that we have visualized the augmented images, let's retrain the model on augmented images.

3.4 Retraining a Model on Augmented Images¶

You noticed that during training, we didn't have to provide the labels. ImageDataGenerator took care of it. As the images are loaded from their directories(cats, dogs), they are augmented and labelled at the same time.

Another thing we can shed light on is the batch size. We have loaded our images in batch size of 20. Usually, the default batch size is 32. The value of the batch size only affect training time. The larger the size, the faster the training, and the smaller the size, the slower the training.

The only issue with the large batch size is that it would requires many steps per epoch in order to give optimal model performance. I have merely used 20 based off the number of images we have in both sets, to facilitate the computation and steps per epochs. But a rule of thumb is to always start with 32. There is this great paper that talks about that: Practical recommendations for gradient-based training of deep architectures - Yoshua Bengio.

To be able to see what the difference data augmentation will make, we will use the same model as before. Let's call it.

In [ ]:

model_2 = classifier()

batch_size = 20

train_steps = 2000/batch_size
val_steps = 1000/batch_size

history_2 = model_2.fit(
      train_generator,
      steps_per_epoch=train_steps,  
      epochs=100,
      validation_data=val_generator,
      validation_steps=val_steps)

model_2 = classifier() batch_size = 20 train_steps = 2000/batch_size val_steps = 1000/batch_size history_2 = model_2.fit( train_generator, steps_per_epoch=train_steps, epochs=100, validation_data=val_generator, validation_steps=val_steps)

Epoch 1/100
100/100 [==============================] - 27s 262ms/step - loss: 0.9125 - accuracy: 0.5150 - val_loss: 0.6847 - val_accuracy: 0.5680
Epoch 2/100
100/100 [==============================] - 26s 260ms/step - loss: 0.7163 - accuracy: 0.5620 - val_loss: 0.7020 - val_accuracy: 0.5400
Epoch 3/100
100/100 [==============================] - 26s 260ms/step - loss: 0.6724 - accuracy: 0.6190 - val_loss: 0.6039 - val_accuracy: 0.6670
Epoch 4/100
100/100 [==============================] - 26s 261ms/step - loss: 0.6502 - accuracy: 0.6325 - val_loss: 0.6332 - val_accuracy: 0.6580
Epoch 5/100
100/100 [==============================] - 26s 261ms/step - loss: 0.6322 - accuracy: 0.6530 - val_loss: 0.5951 - val_accuracy: 0.6880
Epoch 6/100
100/100 [==============================] - 26s 261ms/step - loss: 0.6466 - accuracy: 0.6570 - val_loss: 0.5699 - val_accuracy: 0.6920
Epoch 7/100
100/100 [==============================] - 26s 259ms/step - loss: 0.6135 - accuracy: 0.6610 - val_loss: 0.5892 - val_accuracy: 0.6930
Epoch 8/100
100/100 [==============================] - 26s 260ms/step - loss: 0.6110 - accuracy: 0.6825 - val_loss: 0.7350 - val_accuracy: 0.5930
Epoch 9/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5980 - accuracy: 0.6710 - val_loss: 0.5578 - val_accuracy: 0.7080
Epoch 10/100
100/100 [==============================] - 26s 260ms/step - loss: 0.6059 - accuracy: 0.6830 - val_loss: 0.5752 - val_accuracy: 0.6930
Epoch 11/100
100/100 [==============================] - 26s 259ms/step - loss: 0.5952 - accuracy: 0.6975 - val_loss: 0.5849 - val_accuracy: 0.6880
Epoch 12/100
100/100 [==============================] - 26s 259ms/step - loss: 0.5795 - accuracy: 0.7035 - val_loss: 0.5498 - val_accuracy: 0.7100
Epoch 13/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5825 - accuracy: 0.6955 - val_loss: 0.5590 - val_accuracy: 0.7330
Epoch 14/100
100/100 [==============================] - 26s 259ms/step - loss: 0.5684 - accuracy: 0.6995 - val_loss: 0.5590 - val_accuracy: 0.7390
Epoch 15/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5707 - accuracy: 0.7145 - val_loss: 0.5559 - val_accuracy: 0.7290
Epoch 16/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5704 - accuracy: 0.7110 - val_loss: 0.5219 - val_accuracy: 0.7530
Epoch 17/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5568 - accuracy: 0.7220 - val_loss: 0.5282 - val_accuracy: 0.7450
Epoch 18/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5494 - accuracy: 0.7295 - val_loss: 0.5171 - val_accuracy: 0.7490
Epoch 19/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5493 - accuracy: 0.7305 - val_loss: 0.5150 - val_accuracy: 0.7580
Epoch 20/100
100/100 [==============================] - 26s 261ms/step - loss: 0.5572 - accuracy: 0.7180 - val_loss: 0.5092 - val_accuracy: 0.7490
Epoch 21/100
100/100 [==============================] - 26s 261ms/step - loss: 0.5455 - accuracy: 0.7315 - val_loss: 1.3531 - val_accuracy: 0.5490
Epoch 22/100
100/100 [==============================] - 26s 262ms/step - loss: 0.5559 - accuracy: 0.7190 - val_loss: 0.5160 - val_accuracy: 0.7360
Epoch 23/100
100/100 [==============================] - 26s 261ms/step - loss: 0.5346 - accuracy: 0.7480 - val_loss: 0.4923 - val_accuracy: 0.7530
Epoch 24/100
100/100 [==============================] - 26s 261ms/step - loss: 0.5278 - accuracy: 0.7465 - val_loss: 0.6548 - val_accuracy: 0.7320
Epoch 25/100
100/100 [==============================] - 26s 262ms/step - loss: 0.5528 - accuracy: 0.7300 - val_loss: 0.4998 - val_accuracy: 0.7710
Epoch 26/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5262 - accuracy: 0.7380 - val_loss: 0.4864 - val_accuracy: 0.7750
Epoch 27/100
100/100 [==============================] - 26s 261ms/step - loss: 0.5367 - accuracy: 0.7320 - val_loss: 0.5042 - val_accuracy: 0.7560
Epoch 28/100
100/100 [==============================] - 26s 262ms/step - loss: 0.5273 - accuracy: 0.7375 - val_loss: 0.5322 - val_accuracy: 0.7670
Epoch 29/100
100/100 [==============================] - 26s 261ms/step - loss: 0.5140 - accuracy: 0.7510 - val_loss: 0.4883 - val_accuracy: 0.7650
Epoch 30/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5395 - accuracy: 0.7450 - val_loss: 0.5540 - val_accuracy: 0.7410
Epoch 31/100
100/100 [==============================] - 26s 262ms/step - loss: 0.5169 - accuracy: 0.7460 - val_loss: 0.5006 - val_accuracy: 0.7610
Epoch 32/100
100/100 [==============================] - 26s 260ms/step - loss: 0.5229 - accuracy: 0.7455 - val_loss: 0.5133 - val_accuracy: 0.7400
Epoch 33/100
100/100 [==============================] - 26s 262ms/step - loss: 0.5317 - accuracy: 0.7410 - val_loss: 0.5122 - val_accuracy: 0.7460
Epoch 34/100
100/100 [==============================] - 27s 268ms/step - loss: 0.5261 - accuracy: 0.7400 - val_loss: 0.4857 - val_accuracy: 0.7630
Epoch 35/100
100/100 [==============================] - 26s 264ms/step - loss: 0.5165 - accuracy: 0.7560 - val_loss: 0.4647 - val_accuracy: 0.7790
Epoch 36/100
100/100 [==============================] - 26s 262ms/step - loss: 0.5171 - accuracy: 0.7510 - val_loss: 0.4880 - val_accuracy: 0.7550
Epoch 37/100
100/100 [==============================] - 26s 259ms/step - loss: 0.5121 - accuracy: 0.7465 - val_loss: 0.4775 - val_accuracy: 0.7880
Epoch 38/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4949 - accuracy: 0.7685 - val_loss: 0.4815 - val_accuracy: 0.7780
Epoch 39/100
100/100 [==============================] - 26s 262ms/step - loss: 0.5242 - accuracy: 0.7560 - val_loss: 0.5349 - val_accuracy: 0.7260
Epoch 40/100
100/100 [==============================] - 27s 266ms/step - loss: 0.4927 - accuracy: 0.7655 - val_loss: 0.5462 - val_accuracy: 0.7160
Epoch 41/100
100/100 [==============================] - 26s 265ms/step - loss: 0.5141 - accuracy: 0.7600 - val_loss: 0.5143 - val_accuracy: 0.7670
Epoch 42/100
100/100 [==============================] - 26s 264ms/step - loss: 0.4941 - accuracy: 0.7575 - val_loss: 0.5477 - val_accuracy: 0.7700
Epoch 43/100
100/100 [==============================] - 26s 263ms/step - loss: 0.4901 - accuracy: 0.7760 - val_loss: 0.5021 - val_accuracy: 0.7970
Epoch 44/100
100/100 [==============================] - 26s 261ms/step - loss: 0.5001 - accuracy: 0.7560 - val_loss: 0.5141 - val_accuracy: 0.7810
Epoch 45/100
100/100 [==============================] - 26s 264ms/step - loss: 0.5046 - accuracy: 0.7705 - val_loss: 0.4855 - val_accuracy: 0.7600
Epoch 46/100
100/100 [==============================] - 26s 264ms/step - loss: 0.4864 - accuracy: 0.7770 - val_loss: 0.4695 - val_accuracy: 0.7880
Epoch 47/100
100/100 [==============================] - 26s 264ms/step - loss: 0.4881 - accuracy: 0.7765 - val_loss: 0.5015 - val_accuracy: 0.7590
Epoch 48/100
100/100 [==============================] - 26s 262ms/step - loss: 0.4922 - accuracy: 0.7640 - val_loss: 0.5486 - val_accuracy: 0.7590
Epoch 49/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4942 - accuracy: 0.7670 - val_loss: 0.4959 - val_accuracy: 0.7700
Epoch 50/100
100/100 [==============================] - 26s 261ms/step - loss: 0.5053 - accuracy: 0.7575 - val_loss: 0.6068 - val_accuracy: 0.7170
Epoch 51/100
100/100 [==============================] - 26s 263ms/step - loss: 0.4978 - accuracy: 0.7710 - val_loss: 0.4660 - val_accuracy: 0.7860
Epoch 52/100
100/100 [==============================] - 26s 261ms/step - loss: 0.4995 - accuracy: 0.7675 - val_loss: 0.7096 - val_accuracy: 0.6820
Epoch 53/100
100/100 [==============================] - 26s 262ms/step - loss: 0.5044 - accuracy: 0.7650 - val_loss: 0.4931 - val_accuracy: 0.7730
Epoch 54/100
100/100 [==============================] - 26s 262ms/step - loss: 0.4918 - accuracy: 0.7785 - val_loss: 0.4823 - val_accuracy: 0.7650
Epoch 55/100
100/100 [==============================] - 26s 262ms/step - loss: 0.4842 - accuracy: 0.7690 - val_loss: 0.6578 - val_accuracy: 0.7260
Epoch 56/100
100/100 [==============================] - 26s 263ms/step - loss: 0.4991 - accuracy: 0.7630 - val_loss: 0.4485 - val_accuracy: 0.7990
Epoch 57/100
100/100 [==============================] - 26s 263ms/step - loss: 0.4996 - accuracy: 0.7705 - val_loss: 0.6781 - val_accuracy: 0.7250
Epoch 58/100
100/100 [==============================] - 26s 262ms/step - loss: 0.4799 - accuracy: 0.7820 - val_loss: 0.4580 - val_accuracy: 0.7920
Epoch 59/100
100/100 [==============================] - 26s 263ms/step - loss: 0.4987 - accuracy: 0.7590 - val_loss: 0.4679 - val_accuracy: 0.7710
Epoch 60/100
100/100 [==============================] - 26s 262ms/step - loss: 0.4842 - accuracy: 0.7825 - val_loss: 0.6739 - val_accuracy: 0.6980
Epoch 61/100
100/100 [==============================] - 26s 262ms/step - loss: 0.4849 - accuracy: 0.7745 - val_loss: 0.4635 - val_accuracy: 0.7890
Epoch 62/100
100/100 [==============================] - 26s 264ms/step - loss: 0.4711 - accuracy: 0.7830 - val_loss: 0.4571 - val_accuracy: 0.7850
Epoch 63/100
100/100 [==============================] - 26s 263ms/step - loss: 0.4866 - accuracy: 0.7630 - val_loss: 0.4490 - val_accuracy: 0.7980
Epoch 64/100
100/100 [==============================] - 26s 264ms/step - loss: 0.4751 - accuracy: 0.7800 - val_loss: 0.4483 - val_accuracy: 0.8010
Epoch 65/100
100/100 [==============================] - 27s 268ms/step - loss: 0.4659 - accuracy: 0.7815 - val_loss: 0.4698 - val_accuracy: 0.7900
Epoch 66/100
100/100 [==============================] - 27s 267ms/step - loss: 0.4796 - accuracy: 0.7715 - val_loss: 0.4577 - val_accuracy: 0.7840
Epoch 67/100
100/100 [==============================] - 26s 264ms/step - loss: 0.4696 - accuracy: 0.7845 - val_loss: 0.4937 - val_accuracy: 0.7580
Epoch 68/100
100/100 [==============================] - 26s 262ms/step - loss: 0.4880 - accuracy: 0.7780 - val_loss: 0.8981 - val_accuracy: 0.6680
Epoch 69/100
100/100 [==============================] - 26s 261ms/step - loss: 0.4778 - accuracy: 0.7730 - val_loss: 0.4467 - val_accuracy: 0.7980
Epoch 70/100
100/100 [==============================] - 26s 260ms/step - loss: 0.4663 - accuracy: 0.7800 - val_loss: 0.4448 - val_accuracy: 0.8000
Epoch 71/100
100/100 [==============================] - 26s 260ms/step - loss: 0.4749 - accuracy: 0.7795 - val_loss: 0.4991 - val_accuracy: 0.7830
Epoch 72/100
100/100 [==============================] - 26s 261ms/step - loss: 0.4772 - accuracy: 0.7805 - val_loss: 0.4556 - val_accuracy: 0.7680
Epoch 73/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4586 - accuracy: 0.7780 - val_loss: 0.6288 - val_accuracy: 0.7820
Epoch 74/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4761 - accuracy: 0.7830 - val_loss: 0.4684 - val_accuracy: 0.7880
Epoch 75/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4898 - accuracy: 0.7905 - val_loss: 0.5107 - val_accuracy: 0.7830
Epoch 76/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4798 - accuracy: 0.7770 - val_loss: 0.5129 - val_accuracy: 0.7750
Epoch 77/100
100/100 [==============================] - 26s 258ms/step - loss: 0.4637 - accuracy: 0.7780 - val_loss: 0.5715 - val_accuracy: 0.7570
Epoch 78/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4799 - accuracy: 0.7825 - val_loss: 0.4657 - val_accuracy: 0.7810
Epoch 79/100
100/100 [==============================] - 26s 260ms/step - loss: 0.4726 - accuracy: 0.7975 - val_loss: 0.4772 - val_accuracy: 0.7810
Epoch 80/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4624 - accuracy: 0.7915 - val_loss: 0.4837 - val_accuracy: 0.7640
Epoch 81/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4544 - accuracy: 0.7950 - val_loss: 0.4874 - val_accuracy: 0.7760
Epoch 82/100
100/100 [==============================] - 26s 256ms/step - loss: 0.4725 - accuracy: 0.7995 - val_loss: 0.4473 - val_accuracy: 0.8050
Epoch 83/100
100/100 [==============================] - 26s 256ms/step - loss: 0.4580 - accuracy: 0.7885 - val_loss: 0.6340 - val_accuracy: 0.7330
Epoch 84/100
100/100 [==============================] - 26s 257ms/step - loss: 0.4829 - accuracy: 0.7770 - val_loss: 0.9491 - val_accuracy: 0.6530
Epoch 85/100
100/100 [==============================] - 26s 255ms/step - loss: 0.4757 - accuracy: 0.7720 - val_loss: 0.5391 - val_accuracy: 0.7240
Epoch 86/100
100/100 [==============================] - 25s 255ms/step - loss: 0.4659 - accuracy: 0.7815 - val_loss: 0.4597 - val_accuracy: 0.7820
Epoch 87/100
100/100 [==============================] - 26s 256ms/step - loss: 0.4700 - accuracy: 0.7940 - val_loss: 0.5405 - val_accuracy: 0.7680
Epoch 88/100
100/100 [==============================] - 26s 257ms/step - loss: 0.4764 - accuracy: 0.7755 - val_loss: 0.4748 - val_accuracy: 0.7810
Epoch 89/100
100/100 [==============================] - 26s 257ms/step - loss: 0.4651 - accuracy: 0.8025 - val_loss: 0.5026 - val_accuracy: 0.8060
Epoch 90/100
100/100 [==============================] - 26s 257ms/step - loss: 0.4647 - accuracy: 0.7745 - val_loss: 0.4729 - val_accuracy: 0.8000
Epoch 91/100
100/100 [==============================] - 26s 257ms/step - loss: 0.4742 - accuracy: 0.7975 - val_loss: 0.4547 - val_accuracy: 0.8030
Epoch 92/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4620 - accuracy: 0.7840 - val_loss: 0.5222 - val_accuracy: 0.7700
Epoch 93/100
100/100 [==============================] - 26s 258ms/step - loss: 0.4495 - accuracy: 0.7955 - val_loss: 0.4756 - val_accuracy: 0.8070
Epoch 94/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4627 - accuracy: 0.7865 - val_loss: 0.4523 - val_accuracy: 0.7990
Epoch 95/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4650 - accuracy: 0.7850 - val_loss: 0.5393 - val_accuracy: 0.7630
Epoch 96/100
100/100 [==============================] - 26s 258ms/step - loss: 0.4628 - accuracy: 0.7860 - val_loss: 0.4386 - val_accuracy: 0.8020
Epoch 97/100
100/100 [==============================] - 26s 258ms/step - loss: 0.4665 - accuracy: 0.7955 - val_loss: 0.4350 - val_accuracy: 0.8090
Epoch 98/100
100/100 [==============================] - 26s 259ms/step - loss: 0.4646 - accuracy: 0.7860 - val_loss: 0.4551 - val_accuracy: 0.7960
Epoch 99/100
100/100 [==============================] - 26s 261ms/step - loss: 0.4780 - accuracy: 0.7760 - val_loss: 0.5404 - val_accuracy: 0.7260
Epoch 100/100
100/100 [==============================] - 26s 258ms/step - loss: 0.4772 - accuracy: 0.7750 - val_loss: 0.4480 - val_accuracy: 0.7890

3.5 Visualizing The Model Results¶

In [ ]:

model_history_2 = history_2.history

acc = model_history_2['accuracy']
val_acc = model_history_2['val_accuracy']
loss = model_history_2['loss']
val_loss = model_history_2['val_loss']

epochs = history_2.epoch

model_history_2 = history_2.history acc = model_history_2['accuracy'] val_acc = model_history_2['val_accuracy'] loss = model_history_2['loss'] val_loss = model_history_2['val_loss'] epochs = history_2.epoch

In [ ]:

plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

<Figure size 432x288 with 0 Axes>

This is not excellent, but it's alot better than the results we had without data augmentation. you remember that our model was overfitting, but now although it's not smooth, there is a big improvement.

How to improve the results? One might try to tweak layers and filters. Machine Learning is very experimental. It's rare that the first model will work well. The result is a function of time and experimentation.

So, in this case, we can try pretrained models. Pretrained models are open source models that are built by other engineers(often researchers) and we can use them instead of building a network from scratch.

The technique of reusing a pretrained model into a given(similar) task is called transfer learning. Although this will be covered in the next notebook, let's give it a shot right away.

3.6 Further Improvements: Using Pretrained Models¶

Pretrained models works so well for many problems, without the need of building models from scratch.

Imagine how far you get by standing on the shoulder of the giant! By using powerful models trained on big datasets, the results are pretty impressive.

Let's practice that. For more about pretrained models, check the next notebook.

You can find available pretrained models in Keras on Keras Applications.

In [ ]:

pretrained_base_model = keras.applications.InceptionResNetV2(
    weights='imagenet',
    include_top=False, # Drop imagenet classifier on the top
    input_shape=(180,180,3)

)

pretrained_base_model = keras.applications.InceptionResNetV2( weights='imagenet', include_top=False, # Drop imagenet classifier on the top input_shape=(180,180,3) )

Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/inception_resnet_v2/inception_resnet_v2_weights_tf_dim_ordering_tf_kernels_notop.h5
219062272/219055592 [==============================] - 2s 0us/step
219070464/219055592 [==============================] - 2s 0us/step

We then freeze the pretrained base model to avoid retraining the bottom layers.

In [ ]:

for layer in pretrained_base_model.layers:
  layer.trainable = False

for layer in pretrained_base_model.layers: layer.trainable = False

Let's see the summary of the base model.

In [ ]:

pretrained_base_model.summary()

pretrained_base_model.summary()

Model: "inception_resnet_v2"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 180, 180, 3) 0                                            
__________________________________________________________________________________________________
conv2d_9 (Conv2D)               (None, 89, 89, 32)   864         input_1[0][0]                    
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 89, 89, 32)   96          conv2d_9[0][0]                   
__________________________________________________________________________________________________
activation (Activation)         (None, 89, 89, 32)   0           batch_normalization[0][0]        
__________________________________________________________________________________________________
conv2d_10 (Conv2D)              (None, 87, 87, 32)   9216        activation[0][0]                 
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 87, 87, 32)   96          conv2d_10[0][0]                  
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 87, 87, 32)   0           batch_normalization_1[0][0]      
__________________________________________________________________________________________________
conv2d_11 (Conv2D)              (None, 87, 87, 64)   18432       activation_1[0][0]               
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 87, 87, 64)   192         conv2d_11[0][0]                  
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 87, 87, 64)   0           batch_normalization_2[0][0]      
__________________________________________________________________________________________________
max_pooling2d_9 (MaxPooling2D)  (None, 43, 43, 64)   0           activation_2[0][0]               
__________________________________________________________________________________________________
conv2d_12 (Conv2D)              (None, 43, 43, 80)   5120        max_pooling2d_9[0][0]            
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 43, 43, 80)   240         conv2d_12[0][0]                  
__________________________________________________________________________________________________
activation_3 (Activation)       (None, 43, 43, 80)   0           batch_normalization_3[0][0]      
__________________________________________________________________________________________________
conv2d_13 (Conv2D)              (None, 41, 41, 192)  138240      activation_3[0][0]               
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 41, 41, 192)  576         conv2d_13[0][0]                  
__________________________________________________________________________________________________
activation_4 (Activation)       (None, 41, 41, 192)  0           batch_normalization_4[0][0]      
__________________________________________________________________________________________________
max_pooling2d_10 (MaxPooling2D) (None, 20, 20, 192)  0           activation_4[0][0]               
__________________________________________________________________________________________________
conv2d_17 (Conv2D)              (None, 20, 20, 64)   12288       max_pooling2d_10[0][0]           
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 20, 20, 64)   192         conv2d_17[0][0]                  
__________________________________________________________________________________________________
activation_8 (Activation)       (None, 20, 20, 64)   0           batch_normalization_8[0][0]      
__________________________________________________________________________________________________
conv2d_15 (Conv2D)              (None, 20, 20, 48)   9216        max_pooling2d_10[0][0]           
__________________________________________________________________________________________________
conv2d_18 (Conv2D)              (None, 20, 20, 96)   55296       activation_8[0][0]               
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 20, 20, 48)   144         conv2d_15[0][0]                  
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 20, 20, 96)   288         conv2d_18[0][0]                  
__________________________________________________________________________________________________
activation_6 (Activation)       (None, 20, 20, 48)   0           batch_normalization_6[0][0]      
__________________________________________________________________________________________________
activation_9 (Activation)       (None, 20, 20, 96)   0           batch_normalization_9[0][0]      
__________________________________________________________________________________________________
average_pooling2d (AveragePooli (None, 20, 20, 192)  0           max_pooling2d_10[0][0]           
__________________________________________________________________________________________________
conv2d_14 (Conv2D)              (None, 20, 20, 96)   18432       max_pooling2d_10[0][0]           
__________________________________________________________________________________________________
conv2d_16 (Conv2D)              (None, 20, 20, 64)   76800       activation_6[0][0]               
__________________________________________________________________________________________________
conv2d_19 (Conv2D)              (None, 20, 20, 96)   82944       activation_9[0][0]               
__________________________________________________________________________________________________
conv2d_20 (Conv2D)              (None, 20, 20, 64)   12288       average_pooling2d[0][0]          
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 20, 20, 96)   288         conv2d_14[0][0]                  
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 20, 20, 64)   192         conv2d_16[0][0]                  
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 20, 20, 96)   288         conv2d_19[0][0]                  
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 20, 20, 64)   192         conv2d_20[0][0]                  
__________________________________________________________________________________________________
activation_5 (Activation)       (None, 20, 20, 96)   0           batch_normalization_5[0][0]      
__________________________________________________________________________________________________
activation_7 (Activation)       (None, 20, 20, 64)   0           batch_normalization_7[0][0]      
__________________________________________________________________________________________________
activation_10 (Activation)      (None, 20, 20, 96)   0           batch_normalization_10[0][0]     
__________________________________________________________________________________________________
activation_11 (Activation)      (None, 20, 20, 64)   0           batch_normalization_11[0][0]     
__________________________________________________________________________________________________
mixed_5b (Concatenate)          (None, 20, 20, 320)  0           activation_5[0][0]               
                                                                 activation_7[0][0]               
                                                                 activation_10[0][0]              
                                                                 activation_11[0][0]              
__________________________________________________________________________________________________
conv2d_24 (Conv2D)              (None, 20, 20, 32)   10240       mixed_5b[0][0]                   
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 20, 20, 32)   96          conv2d_24[0][0]                  
__________________________________________________________________________________________________
activation_15 (Activation)      (None, 20, 20, 32)   0           batch_normalization_15[0][0]     
__________________________________________________________________________________________________
conv2d_22 (Conv2D)              (None, 20, 20, 32)   10240       mixed_5b[0][0]                   
__________________________________________________________________________________________________
conv2d_25 (Conv2D)              (None, 20, 20, 48)   13824       activation_15[0][0]              
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 20, 20, 32)   96          conv2d_22[0][0]                  
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 20, 20, 48)   144         conv2d_25[0][0]                  
__________________________________________________________________________________________________
activation_13 (Activation)      (None, 20, 20, 32)   0           batch_normalization_13[0][0]     
__________________________________________________________________________________________________
activation_16 (Activation)      (None, 20, 20, 48)   0           batch_normalization_16[0][0]     
__________________________________________________________________________________________________
conv2d_21 (Conv2D)              (None, 20, 20, 32)   10240       mixed_5b[0][0]                   
__________________________________________________________________________________________________
conv2d_23 (Conv2D)              (None, 20, 20, 32)   9216        activation_13[0][0]              
__________________________________________________________________________________________________
conv2d_26 (Conv2D)              (None, 20, 20, 64)   27648       activation_16[0][0]              
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 20, 20, 32)   96          conv2d_21[0][0]                  
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 20, 20, 32)   96          conv2d_23[0][0]                  
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 20, 20, 64)   192         conv2d_26[0][0]                  
__________________________________________________________________________________________________
activation_12 (Activation)      (None, 20, 20, 32)   0           batch_normalization_12[0][0]     
__________________________________________________________________________________________________
activation_14 (Activation)      (None, 20, 20, 32)   0           batch_normalization_14[0][0]     
__________________________________________________________________________________________________
activation_17 (Activation)      (None, 20, 20, 64)   0           batch_normalization_17[0][0]     
__________________________________________________________________________________________________
block35_1_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_12[0][0]              
                                                                 activation_14[0][0]              
                                                                 activation_17[0][0]              
__________________________________________________________________________________________________
block35_1_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_1_mixed[0][0]            
__________________________________________________________________________________________________
block35_1 (Lambda)              (None, 20, 20, 320)  0           mixed_5b[0][0]                   
                                                                 block35_1_conv[0][0]             
__________________________________________________________________________________________________
block35_1_ac (Activation)       (None, 20, 20, 320)  0           block35_1[0][0]                  
__________________________________________________________________________________________________
conv2d_30 (Conv2D)              (None, 20, 20, 32)   10240       block35_1_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_21 (BatchNo (None, 20, 20, 32)   96          conv2d_30[0][0]                  
__________________________________________________________________________________________________
activation_21 (Activation)      (None, 20, 20, 32)   0           batch_normalization_21[0][0]     
__________________________________________________________________________________________________
conv2d_28 (Conv2D)              (None, 20, 20, 32)   10240       block35_1_ac[0][0]               
__________________________________________________________________________________________________
conv2d_31 (Conv2D)              (None, 20, 20, 48)   13824       activation_21[0][0]              
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 20, 20, 32)   96          conv2d_28[0][0]                  
__________________________________________________________________________________________________
batch_normalization_22 (BatchNo (None, 20, 20, 48)   144         conv2d_31[0][0]                  
__________________________________________________________________________________________________
activation_19 (Activation)      (None, 20, 20, 32)   0           batch_normalization_19[0][0]     
__________________________________________________________________________________________________
activation_22 (Activation)      (None, 20, 20, 48)   0           batch_normalization_22[0][0]     
__________________________________________________________________________________________________
conv2d_27 (Conv2D)              (None, 20, 20, 32)   10240       block35_1_ac[0][0]               
__________________________________________________________________________________________________
conv2d_29 (Conv2D)              (None, 20, 20, 32)   9216        activation_19[0][0]              
__________________________________________________________________________________________________
conv2d_32 (Conv2D)              (None, 20, 20, 64)   27648       activation_22[0][0]              
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 20, 20, 32)   96          conv2d_27[0][0]                  
__________________________________________________________________________________________________
batch_normalization_20 (BatchNo (None, 20, 20, 32)   96          conv2d_29[0][0]                  
__________________________________________________________________________________________________
batch_normalization_23 (BatchNo (None, 20, 20, 64)   192         conv2d_32[0][0]                  
__________________________________________________________________________________________________
activation_18 (Activation)      (None, 20, 20, 32)   0           batch_normalization_18[0][0]     
__________________________________________________________________________________________________
activation_20 (Activation)      (None, 20, 20, 32)   0           batch_normalization_20[0][0]     
__________________________________________________________________________________________________
activation_23 (Activation)      (None, 20, 20, 64)   0           batch_normalization_23[0][0]     
__________________________________________________________________________________________________
block35_2_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_18[0][0]              
                                                                 activation_20[0][0]              
                                                                 activation_23[0][0]              
__________________________________________________________________________________________________
block35_2_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_2_mixed[0][0]            
__________________________________________________________________________________________________
block35_2 (Lambda)              (None, 20, 20, 320)  0           block35_1_ac[0][0]               
                                                                 block35_2_conv[0][0]             
__________________________________________________________________________________________________
block35_2_ac (Activation)       (None, 20, 20, 320)  0           block35_2[0][0]                  
__________________________________________________________________________________________________
conv2d_36 (Conv2D)              (None, 20, 20, 32)   10240       block35_2_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_27 (BatchNo (None, 20, 20, 32)   96          conv2d_36[0][0]                  
__________________________________________________________________________________________________
activation_27 (Activation)      (None, 20, 20, 32)   0           batch_normalization_27[0][0]     
__________________________________________________________________________________________________
conv2d_34 (Conv2D)              (None, 20, 20, 32)   10240       block35_2_ac[0][0]               
__________________________________________________________________________________________________
conv2d_37 (Conv2D)              (None, 20, 20, 48)   13824       activation_27[0][0]              
__________________________________________________________________________________________________
batch_normalization_25 (BatchNo (None, 20, 20, 32)   96          conv2d_34[0][0]                  
__________________________________________________________________________________________________
batch_normalization_28 (BatchNo (None, 20, 20, 48)   144         conv2d_37[0][0]                  
__________________________________________________________________________________________________
activation_25 (Activation)      (None, 20, 20, 32)   0           batch_normalization_25[0][0]     
__________________________________________________________________________________________________
activation_28 (Activation)      (None, 20, 20, 48)   0           batch_normalization_28[0][0]     
__________________________________________________________________________________________________
conv2d_33 (Conv2D)              (None, 20, 20, 32)   10240       block35_2_ac[0][0]               
__________________________________________________________________________________________________
conv2d_35 (Conv2D)              (None, 20, 20, 32)   9216        activation_25[0][0]              
__________________________________________________________________________________________________
conv2d_38 (Conv2D)              (None, 20, 20, 64)   27648       activation_28[0][0]              
__________________________________________________________________________________________________
batch_normalization_24 (BatchNo (None, 20, 20, 32)   96          conv2d_33[0][0]                  
__________________________________________________________________________________________________
batch_normalization_26 (BatchNo (None, 20, 20, 32)   96          conv2d_35[0][0]                  
__________________________________________________________________________________________________
batch_normalization_29 (BatchNo (None, 20, 20, 64)   192         conv2d_38[0][0]                  
__________________________________________________________________________________________________
activation_24 (Activation)      (None, 20, 20, 32)   0           batch_normalization_24[0][0]     
__________________________________________________________________________________________________
activation_26 (Activation)      (None, 20, 20, 32)   0           batch_normalization_26[0][0]     
__________________________________________________________________________________________________
activation_29 (Activation)      (None, 20, 20, 64)   0           batch_normalization_29[0][0]     
__________________________________________________________________________________________________
block35_3_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_24[0][0]              
                                                                 activation_26[0][0]              
                                                                 activation_29[0][0]              
__________________________________________________________________________________________________
block35_3_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_3_mixed[0][0]            
__________________________________________________________________________________________________
block35_3 (Lambda)              (None, 20, 20, 320)  0           block35_2_ac[0][0]               
                                                                 block35_3_conv[0][0]             
__________________________________________________________________________________________________
block35_3_ac (Activation)       (None, 20, 20, 320)  0           block35_3[0][0]                  
__________________________________________________________________________________________________
conv2d_42 (Conv2D)              (None, 20, 20, 32)   10240       block35_3_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_33 (BatchNo (None, 20, 20, 32)   96          conv2d_42[0][0]                  
__________________________________________________________________________________________________
activation_33 (Activation)      (None, 20, 20, 32)   0           batch_normalization_33[0][0]     
__________________________________________________________________________________________________
conv2d_40 (Conv2D)              (None, 20, 20, 32)   10240       block35_3_ac[0][0]               
__________________________________________________________________________________________________
conv2d_43 (Conv2D)              (None, 20, 20, 48)   13824       activation_33[0][0]              
__________________________________________________________________________________________________
batch_normalization_31 (BatchNo (None, 20, 20, 32)   96          conv2d_40[0][0]                  
__________________________________________________________________________________________________
batch_normalization_34 (BatchNo (None, 20, 20, 48)   144         conv2d_43[0][0]                  
__________________________________________________________________________________________________
activation_31 (Activation)      (None, 20, 20, 32)   0           batch_normalization_31[0][0]     
__________________________________________________________________________________________________
activation_34 (Activation)      (None, 20, 20, 48)   0           batch_normalization_34[0][0]     
__________________________________________________________________________________________________
conv2d_39 (Conv2D)              (None, 20, 20, 32)   10240       block35_3_ac[0][0]               
__________________________________________________________________________________________________
conv2d_41 (Conv2D)              (None, 20, 20, 32)   9216        activation_31[0][0]              
__________________________________________________________________________________________________
conv2d_44 (Conv2D)              (None, 20, 20, 64)   27648       activation_34[0][0]              
__________________________________________________________________________________________________
batch_normalization_30 (BatchNo (None, 20, 20, 32)   96          conv2d_39[0][0]                  
__________________________________________________________________________________________________
batch_normalization_32 (BatchNo (None, 20, 20, 32)   96          conv2d_41[0][0]                  
__________________________________________________________________________________________________
batch_normalization_35 (BatchNo (None, 20, 20, 64)   192         conv2d_44[0][0]                  
__________________________________________________________________________________________________
activation_30 (Activation)      (None, 20, 20, 32)   0           batch_normalization_30[0][0]     
__________________________________________________________________________________________________
activation_32 (Activation)      (None, 20, 20, 32)   0           batch_normalization_32[0][0]     
__________________________________________________________________________________________________
activation_35 (Activation)      (None, 20, 20, 64)   0           batch_normalization_35[0][0]     
__________________________________________________________________________________________________
block35_4_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_30[0][0]              
                                                                 activation_32[0][0]              
                                                                 activation_35[0][0]              
__________________________________________________________________________________________________
block35_4_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_4_mixed[0][0]            
__________________________________________________________________________________________________
block35_4 (Lambda)              (None, 20, 20, 320)  0           block35_3_ac[0][0]               
                                                                 block35_4_conv[0][0]             
__________________________________________________________________________________________________
block35_4_ac (Activation)       (None, 20, 20, 320)  0           block35_4[0][0]                  
__________________________________________________________________________________________________
conv2d_48 (Conv2D)              (None, 20, 20, 32)   10240       block35_4_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_39 (BatchNo (None, 20, 20, 32)   96          conv2d_48[0][0]                  
__________________________________________________________________________________________________
activation_39 (Activation)      (None, 20, 20, 32)   0           batch_normalization_39[0][0]     
__________________________________________________________________________________________________
conv2d_46 (Conv2D)              (None, 20, 20, 32)   10240       block35_4_ac[0][0]               
__________________________________________________________________________________________________
conv2d_49 (Conv2D)              (None, 20, 20, 48)   13824       activation_39[0][0]              
__________________________________________________________________________________________________
batch_normalization_37 (BatchNo (None, 20, 20, 32)   96          conv2d_46[0][0]                  
__________________________________________________________________________________________________
batch_normalization_40 (BatchNo (None, 20, 20, 48)   144         conv2d_49[0][0]                  
__________________________________________________________________________________________________
activation_37 (Activation)      (None, 20, 20, 32)   0           batch_normalization_37[0][0]     
__________________________________________________________________________________________________
activation_40 (Activation)      (None, 20, 20, 48)   0           batch_normalization_40[0][0]     
__________________________________________________________________________________________________
conv2d_45 (Conv2D)              (None, 20, 20, 32)   10240       block35_4_ac[0][0]               
__________________________________________________________________________________________________
conv2d_47 (Conv2D)              (None, 20, 20, 32)   9216        activation_37[0][0]              
__________________________________________________________________________________________________
conv2d_50 (Conv2D)              (None, 20, 20, 64)   27648       activation_40[0][0]              
__________________________________________________________________________________________________
batch_normalization_36 (BatchNo (None, 20, 20, 32)   96          conv2d_45[0][0]                  
__________________________________________________________________________________________________
batch_normalization_38 (BatchNo (None, 20, 20, 32)   96          conv2d_47[0][0]                  
__________________________________________________________________________________________________
batch_normalization_41 (BatchNo (None, 20, 20, 64)   192         conv2d_50[0][0]                  
__________________________________________________________________________________________________
activation_36 (Activation)      (None, 20, 20, 32)   0           batch_normalization_36[0][0]     
__________________________________________________________________________________________________
activation_38 (Activation)      (None, 20, 20, 32)   0           batch_normalization_38[0][0]     
__________________________________________________________________________________________________
activation_41 (Activation)      (None, 20, 20, 64)   0           batch_normalization_41[0][0]     
__________________________________________________________________________________________________
block35_5_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_36[0][0]              
                                                                 activation_38[0][0]              
                                                                 activation_41[0][0]              
__________________________________________________________________________________________________
block35_5_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_5_mixed[0][0]            
__________________________________________________________________________________________________
block35_5 (Lambda)              (None, 20, 20, 320)  0           block35_4_ac[0][0]               
                                                                 block35_5_conv[0][0]             
__________________________________________________________________________________________________
block35_5_ac (Activation)       (None, 20, 20, 320)  0           block35_5[0][0]                  
__________________________________________________________________________________________________
conv2d_54 (Conv2D)              (None, 20, 20, 32)   10240       block35_5_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_45 (BatchNo (None, 20, 20, 32)   96          conv2d_54[0][0]                  
__________________________________________________________________________________________________
activation_45 (Activation)      (None, 20, 20, 32)   0           batch_normalization_45[0][0]     
__________________________________________________________________________________________________
conv2d_52 (Conv2D)              (None, 20, 20, 32)   10240       block35_5_ac[0][0]               
__________________________________________________________________________________________________
conv2d_55 (Conv2D)              (None, 20, 20, 48)   13824       activation_45[0][0]              
__________________________________________________________________________________________________
batch_normalization_43 (BatchNo (None, 20, 20, 32)   96          conv2d_52[0][0]                  
__________________________________________________________________________________________________
batch_normalization_46 (BatchNo (None, 20, 20, 48)   144         conv2d_55[0][0]                  
__________________________________________________________________________________________________
activation_43 (Activation)      (None, 20, 20, 32)   0           batch_normalization_43[0][0]     
__________________________________________________________________________________________________
activation_46 (Activation)      (None, 20, 20, 48)   0           batch_normalization_46[0][0]     
__________________________________________________________________________________________________
conv2d_51 (Conv2D)              (None, 20, 20, 32)   10240       block35_5_ac[0][0]               
__________________________________________________________________________________________________
conv2d_53 (Conv2D)              (None, 20, 20, 32)   9216        activation_43[0][0]              
__________________________________________________________________________________________________
conv2d_56 (Conv2D)              (None, 20, 20, 64)   27648       activation_46[0][0]              
__________________________________________________________________________________________________
batch_normalization_42 (BatchNo (None, 20, 20, 32)   96          conv2d_51[0][0]                  
__________________________________________________________________________________________________
batch_normalization_44 (BatchNo (None, 20, 20, 32)   96          conv2d_53[0][0]                  
__________________________________________________________________________________________________
batch_normalization_47 (BatchNo (None, 20, 20, 64)   192         conv2d_56[0][0]                  
__________________________________________________________________________________________________
activation_42 (Activation)      (None, 20, 20, 32)   0           batch_normalization_42[0][0]     
__________________________________________________________________________________________________
activation_44 (Activation)      (None, 20, 20, 32)   0           batch_normalization_44[0][0]     
__________________________________________________________________________________________________
activation_47 (Activation)      (None, 20, 20, 64)   0           batch_normalization_47[0][0]     
__________________________________________________________________________________________________
block35_6_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_42[0][0]              
                                                                 activation_44[0][0]              
                                                                 activation_47[0][0]              
__________________________________________________________________________________________________
block35_6_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_6_mixed[0][0]            
__________________________________________________________________________________________________
block35_6 (Lambda)              (None, 20, 20, 320)  0           block35_5_ac[0][0]               
                                                                 block35_6_conv[0][0]             
__________________________________________________________________________________________________
block35_6_ac (Activation)       (None, 20, 20, 320)  0           block35_6[0][0]                  
__________________________________________________________________________________________________
conv2d_60 (Conv2D)              (None, 20, 20, 32)   10240       block35_6_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_51 (BatchNo (None, 20, 20, 32)   96          conv2d_60[0][0]                  
__________________________________________________________________________________________________
activation_51 (Activation)      (None, 20, 20, 32)   0           batch_normalization_51[0][0]     
__________________________________________________________________________________________________
conv2d_58 (Conv2D)              (None, 20, 20, 32)   10240       block35_6_ac[0][0]               
__________________________________________________________________________________________________
conv2d_61 (Conv2D)              (None, 20, 20, 48)   13824       activation_51[0][0]              
__________________________________________________________________________________________________
batch_normalization_49 (BatchNo (None, 20, 20, 32)   96          conv2d_58[0][0]                  
__________________________________________________________________________________________________
batch_normalization_52 (BatchNo (None, 20, 20, 48)   144         conv2d_61[0][0]                  
__________________________________________________________________________________________________
activation_49 (Activation)      (None, 20, 20, 32)   0           batch_normalization_49[0][0]     
__________________________________________________________________________________________________
activation_52 (Activation)      (None, 20, 20, 48)   0           batch_normalization_52[0][0]     
__________________________________________________________________________________________________
conv2d_57 (Conv2D)              (None, 20, 20, 32)   10240       block35_6_ac[0][0]               
__________________________________________________________________________________________________
conv2d_59 (Conv2D)              (None, 20, 20, 32)   9216        activation_49[0][0]              
__________________________________________________________________________________________________
conv2d_62 (Conv2D)              (None, 20, 20, 64)   27648       activation_52[0][0]              
__________________________________________________________________________________________________
batch_normalization_48 (BatchNo (None, 20, 20, 32)   96          conv2d_57[0][0]                  
__________________________________________________________________________________________________
batch_normalization_50 (BatchNo (None, 20, 20, 32)   96          conv2d_59[0][0]                  
__________________________________________________________________________________________________
batch_normalization_53 (BatchNo (None, 20, 20, 64)   192         conv2d_62[0][0]                  
__________________________________________________________________________________________________
activation_48 (Activation)      (None, 20, 20, 32)   0           batch_normalization_48[0][0]     
__________________________________________________________________________________________________
activation_50 (Activation)      (None, 20, 20, 32)   0           batch_normalization_50[0][0]     
__________________________________________________________________________________________________
activation_53 (Activation)      (None, 20, 20, 64)   0           batch_normalization_53[0][0]     
__________________________________________________________________________________________________
block35_7_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_48[0][0]              
                                                                 activation_50[0][0]              
                                                                 activation_53[0][0]              
__________________________________________________________________________________________________
block35_7_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_7_mixed[0][0]            
__________________________________________________________________________________________________
block35_7 (Lambda)              (None, 20, 20, 320)  0           block35_6_ac[0][0]               
                                                                 block35_7_conv[0][0]             
__________________________________________________________________________________________________
block35_7_ac (Activation)       (None, 20, 20, 320)  0           block35_7[0][0]                  
__________________________________________________________________________________________________
conv2d_66 (Conv2D)              (None, 20, 20, 32)   10240       block35_7_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_57 (BatchNo (None, 20, 20, 32)   96          conv2d_66[0][0]                  
__________________________________________________________________________________________________
activation_57 (Activation)      (None, 20, 20, 32)   0           batch_normalization_57[0][0]     
__________________________________________________________________________________________________
conv2d_64 (Conv2D)              (None, 20, 20, 32)   10240       block35_7_ac[0][0]               
__________________________________________________________________________________________________
conv2d_67 (Conv2D)              (None, 20, 20, 48)   13824       activation_57[0][0]              
__________________________________________________________________________________________________
batch_normalization_55 (BatchNo (None, 20, 20, 32)   96          conv2d_64[0][0]                  
__________________________________________________________________________________________________
batch_normalization_58 (BatchNo (None, 20, 20, 48)   144         conv2d_67[0][0]                  
__________________________________________________________________________________________________
activation_55 (Activation)      (None, 20, 20, 32)   0           batch_normalization_55[0][0]     
__________________________________________________________________________________________________
activation_58 (Activation)      (None, 20, 20, 48)   0           batch_normalization_58[0][0]     
__________________________________________________________________________________________________
conv2d_63 (Conv2D)              (None, 20, 20, 32)   10240       block35_7_ac[0][0]               
__________________________________________________________________________________________________
conv2d_65 (Conv2D)              (None, 20, 20, 32)   9216        activation_55[0][0]              
__________________________________________________________________________________________________
conv2d_68 (Conv2D)              (None, 20, 20, 64)   27648       activation_58[0][0]              
__________________________________________________________________________________________________
batch_normalization_54 (BatchNo (None, 20, 20, 32)   96          conv2d_63[0][0]                  
__________________________________________________________________________________________________
batch_normalization_56 (BatchNo (None, 20, 20, 32)   96          conv2d_65[0][0]                  
__________________________________________________________________________________________________
batch_normalization_59 (BatchNo (None, 20, 20, 64)   192         conv2d_68[0][0]                  
__________________________________________________________________________________________________
activation_54 (Activation)      (None, 20, 20, 32)   0           batch_normalization_54[0][0]     
__________________________________________________________________________________________________
activation_56 (Activation)      (None, 20, 20, 32)   0           batch_normalization_56[0][0]     
__________________________________________________________________________________________________
activation_59 (Activation)      (None, 20, 20, 64)   0           batch_normalization_59[0][0]     
__________________________________________________________________________________________________
block35_8_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_54[0][0]              
                                                                 activation_56[0][0]              
                                                                 activation_59[0][0]              
__________________________________________________________________________________________________
block35_8_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_8_mixed[0][0]            
__________________________________________________________________________________________________
block35_8 (Lambda)              (None, 20, 20, 320)  0           block35_7_ac[0][0]               
                                                                 block35_8_conv[0][0]             
__________________________________________________________________________________________________
block35_8_ac (Activation)       (None, 20, 20, 320)  0           block35_8[0][0]                  
__________________________________________________________________________________________________
conv2d_72 (Conv2D)              (None, 20, 20, 32)   10240       block35_8_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_63 (BatchNo (None, 20, 20, 32)   96          conv2d_72[0][0]                  
__________________________________________________________________________________________________
activation_63 (Activation)      (None, 20, 20, 32)   0           batch_normalization_63[0][0]     
__________________________________________________________________________________________________
conv2d_70 (Conv2D)              (None, 20, 20, 32)   10240       block35_8_ac[0][0]               
__________________________________________________________________________________________________
conv2d_73 (Conv2D)              (None, 20, 20, 48)   13824       activation_63[0][0]              
__________________________________________________________________________________________________
batch_normalization_61 (BatchNo (None, 20, 20, 32)   96          conv2d_70[0][0]                  
__________________________________________________________________________________________________
batch_normalization_64 (BatchNo (None, 20, 20, 48)   144         conv2d_73[0][0]                  
__________________________________________________________________________________________________
activation_61 (Activation)      (None, 20, 20, 32)   0           batch_normalization_61[0][0]     
__________________________________________________________________________________________________
activation_64 (Activation)      (None, 20, 20, 48)   0           batch_normalization_64[0][0]     
__________________________________________________________________________________________________
conv2d_69 (Conv2D)              (None, 20, 20, 32)   10240       block35_8_ac[0][0]               
__________________________________________________________________________________________________
conv2d_71 (Conv2D)              (None, 20, 20, 32)   9216        activation_61[0][0]              
__________________________________________________________________________________________________
conv2d_74 (Conv2D)              (None, 20, 20, 64)   27648       activation_64[0][0]              
__________________________________________________________________________________________________
batch_normalization_60 (BatchNo (None, 20, 20, 32)   96          conv2d_69[0][0]                  
__________________________________________________________________________________________________
batch_normalization_62 (BatchNo (None, 20, 20, 32)   96          conv2d_71[0][0]                  
__________________________________________________________________________________________________
batch_normalization_65 (BatchNo (None, 20, 20, 64)   192         conv2d_74[0][0]                  
__________________________________________________________________________________________________
activation_60 (Activation)      (None, 20, 20, 32)   0           batch_normalization_60[0][0]     
__________________________________________________________________________________________________
activation_62 (Activation)      (None, 20, 20, 32)   0           batch_normalization_62[0][0]     
__________________________________________________________________________________________________
activation_65 (Activation)      (None, 20, 20, 64)   0           batch_normalization_65[0][0]     
__________________________________________________________________________________________________
block35_9_mixed (Concatenate)   (None, 20, 20, 128)  0           activation_60[0][0]              
                                                                 activation_62[0][0]              
                                                                 activation_65[0][0]              
__________________________________________________________________________________________________
block35_9_conv (Conv2D)         (None, 20, 20, 320)  41280       block35_9_mixed[0][0]            
__________________________________________________________________________________________________
block35_9 (Lambda)              (None, 20, 20, 320)  0           block35_8_ac[0][0]               
                                                                 block35_9_conv[0][0]             
__________________________________________________________________________________________________
block35_9_ac (Activation)       (None, 20, 20, 320)  0           block35_9[0][0]                  
__________________________________________________________________________________________________
conv2d_78 (Conv2D)              (None, 20, 20, 32)   10240       block35_9_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_69 (BatchNo (None, 20, 20, 32)   96          conv2d_78[0][0]                  
__________________________________________________________________________________________________
activation_69 (Activation)      (None, 20, 20, 32)   0           batch_normalization_69[0][0]     
__________________________________________________________________________________________________
conv2d_76 (Conv2D)              (None, 20, 20, 32)   10240       block35_9_ac[0][0]               
__________________________________________________________________________________________________
conv2d_79 (Conv2D)              (None, 20, 20, 48)   13824       activation_69[0][0]              
__________________________________________________________________________________________________
batch_normalization_67 (BatchNo (None, 20, 20, 32)   96          conv2d_76[0][0]                  
__________________________________________________________________________________________________
batch_normalization_70 (BatchNo (None, 20, 20, 48)   144         conv2d_79[0][0]                  
__________________________________________________________________________________________________
activation_67 (Activation)      (None, 20, 20, 32)   0           batch_normalization_67[0][0]     
__________________________________________________________________________________________________
activation_70 (Activation)      (None, 20, 20, 48)   0           batch_normalization_70[0][0]     
__________________________________________________________________________________________________
conv2d_75 (Conv2D)              (None, 20, 20, 32)   10240       block35_9_ac[0][0]               
__________________________________________________________________________________________________
conv2d_77 (Conv2D)              (None, 20, 20, 32)   9216        activation_67[0][0]              
__________________________________________________________________________________________________
conv2d_80 (Conv2D)              (None, 20, 20, 64)   27648       activation_70[0][0]              
__________________________________________________________________________________________________
batch_normalization_66 (BatchNo (None, 20, 20, 32)   96          conv2d_75[0][0]                  
__________________________________________________________________________________________________
batch_normalization_68 (BatchNo (None, 20, 20, 32)   96          conv2d_77[0][0]                  
__________________________________________________________________________________________________
batch_normalization_71 (BatchNo (None, 20, 20, 64)   192         conv2d_80[0][0]                  
__________________________________________________________________________________________________
activation_66 (Activation)      (None, 20, 20, 32)   0           batch_normalization_66[0][0]     
__________________________________________________________________________________________________
activation_68 (Activation)      (None, 20, 20, 32)   0           batch_normalization_68[0][0]     
__________________________________________________________________________________________________
activation_71 (Activation)      (None, 20, 20, 64)   0           batch_normalization_71[0][0]     
__________________________________________________________________________________________________
block35_10_mixed (Concatenate)  (None, 20, 20, 128)  0           activation_66[0][0]              
                                                                 activation_68[0][0]              
                                                                 activation_71[0][0]              
__________________________________________________________________________________________________
block35_10_conv (Conv2D)        (None, 20, 20, 320)  41280       block35_10_mixed[0][0]           
__________________________________________________________________________________________________
block35_10 (Lambda)             (None, 20, 20, 320)  0           block35_9_ac[0][0]               
                                                                 block35_10_conv[0][0]            
__________________________________________________________________________________________________
block35_10_ac (Activation)      (None, 20, 20, 320)  0           block35_10[0][0]                 
__________________________________________________________________________________________________
conv2d_82 (Conv2D)              (None, 20, 20, 256)  81920       block35_10_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_73 (BatchNo (None, 20, 20, 256)  768         conv2d_82[0][0]                  
__________________________________________________________________________________________________
activation_73 (Activation)      (None, 20, 20, 256)  0           batch_normalization_73[0][0]     
__________________________________________________________________________________________________
conv2d_83 (Conv2D)              (None, 20, 20, 256)  589824      activation_73[0][0]              
__________________________________________________________________________________________________
batch_normalization_74 (BatchNo (None, 20, 20, 256)  768         conv2d_83[0][0]                  
__________________________________________________________________________________________________
activation_74 (Activation)      (None, 20, 20, 256)  0           batch_normalization_74[0][0]     
__________________________________________________________________________________________________
conv2d_81 (Conv2D)              (None, 9, 9, 384)    1105920     block35_10_ac[0][0]              
__________________________________________________________________________________________________
conv2d_84 (Conv2D)              (None, 9, 9, 384)    884736      activation_74[0][0]              
__________________________________________________________________________________________________
batch_normalization_72 (BatchNo (None, 9, 9, 384)    1152        conv2d_81[0][0]                  
__________________________________________________________________________________________________
batch_normalization_75 (BatchNo (None, 9, 9, 384)    1152        conv2d_84[0][0]                  
__________________________________________________________________________________________________
activation_72 (Activation)      (None, 9, 9, 384)    0           batch_normalization_72[0][0]     
__________________________________________________________________________________________________
activation_75 (Activation)      (None, 9, 9, 384)    0           batch_normalization_75[0][0]     
__________________________________________________________________________________________________
max_pooling2d_11 (MaxPooling2D) (None, 9, 9, 320)    0           block35_10_ac[0][0]              
__________________________________________________________________________________________________
mixed_6a (Concatenate)          (None, 9, 9, 1088)   0           activation_72[0][0]              
                                                                 activation_75[0][0]              
                                                                 max_pooling2d_11[0][0]           
__________________________________________________________________________________________________
conv2d_86 (Conv2D)              (None, 9, 9, 128)    139264      mixed_6a[0][0]                   
__________________________________________________________________________________________________
batch_normalization_77 (BatchNo (None, 9, 9, 128)    384         conv2d_86[0][0]                  
__________________________________________________________________________________________________
activation_77 (Activation)      (None, 9, 9, 128)    0           batch_normalization_77[0][0]     
__________________________________________________________________________________________________
conv2d_87 (Conv2D)              (None, 9, 9, 160)    143360      activation_77[0][0]              
__________________________________________________________________________________________________
batch_normalization_78 (BatchNo (None, 9, 9, 160)    480         conv2d_87[0][0]                  
__________________________________________________________________________________________________
activation_78 (Activation)      (None, 9, 9, 160)    0           batch_normalization_78[0][0]     
__________________________________________________________________________________________________
conv2d_85 (Conv2D)              (None, 9, 9, 192)    208896      mixed_6a[0][0]                   
__________________________________________________________________________________________________
conv2d_88 (Conv2D)              (None, 9, 9, 192)    215040      activation_78[0][0]              
__________________________________________________________________________________________________
batch_normalization_76 (BatchNo (None, 9, 9, 192)    576         conv2d_85[0][0]                  
__________________________________________________________________________________________________
batch_normalization_79 (BatchNo (None, 9, 9, 192)    576         conv2d_88[0][0]                  
__________________________________________________________________________________________________
activation_76 (Activation)      (None, 9, 9, 192)    0           batch_normalization_76[0][0]     
__________________________________________________________________________________________________
activation_79 (Activation)      (None, 9, 9, 192)    0           batch_normalization_79[0][0]     
__________________________________________________________________________________________________
block17_1_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_76[0][0]              
                                                                 activation_79[0][0]              
__________________________________________________________________________________________________
block17_1_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_1_mixed[0][0]            
__________________________________________________________________________________________________
block17_1 (Lambda)              (None, 9, 9, 1088)   0           mixed_6a[0][0]                   
                                                                 block17_1_conv[0][0]             
__________________________________________________________________________________________________
block17_1_ac (Activation)       (None, 9, 9, 1088)   0           block17_1[0][0]                  
__________________________________________________________________________________________________
conv2d_90 (Conv2D)              (None, 9, 9, 128)    139264      block17_1_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_81 (BatchNo (None, 9, 9, 128)    384         conv2d_90[0][0]                  
__________________________________________________________________________________________________
activation_81 (Activation)      (None, 9, 9, 128)    0           batch_normalization_81[0][0]     
__________________________________________________________________________________________________
conv2d_91 (Conv2D)              (None, 9, 9, 160)    143360      activation_81[0][0]              
__________________________________________________________________________________________________
batch_normalization_82 (BatchNo (None, 9, 9, 160)    480         conv2d_91[0][0]                  
__________________________________________________________________________________________________
activation_82 (Activation)      (None, 9, 9, 160)    0           batch_normalization_82[0][0]     
__________________________________________________________________________________________________
conv2d_89 (Conv2D)              (None, 9, 9, 192)    208896      block17_1_ac[0][0]               
__________________________________________________________________________________________________
conv2d_92 (Conv2D)              (None, 9, 9, 192)    215040      activation_82[0][0]              
__________________________________________________________________________________________________
batch_normalization_80 (BatchNo (None, 9, 9, 192)    576         conv2d_89[0][0]                  
__________________________________________________________________________________________________
batch_normalization_83 (BatchNo (None, 9, 9, 192)    576         conv2d_92[0][0]                  
__________________________________________________________________________________________________
activation_80 (Activation)      (None, 9, 9, 192)    0           batch_normalization_80[0][0]     
__________________________________________________________________________________________________
activation_83 (Activation)      (None, 9, 9, 192)    0           batch_normalization_83[0][0]     
__________________________________________________________________________________________________
block17_2_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_80[0][0]              
                                                                 activation_83[0][0]              
__________________________________________________________________________________________________
block17_2_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_2_mixed[0][0]            
__________________________________________________________________________________________________
block17_2 (Lambda)              (None, 9, 9, 1088)   0           block17_1_ac[0][0]               
                                                                 block17_2_conv[0][0]             
__________________________________________________________________________________________________
block17_2_ac (Activation)       (None, 9, 9, 1088)   0           block17_2[0][0]                  
__________________________________________________________________________________________________
conv2d_94 (Conv2D)              (None, 9, 9, 128)    139264      block17_2_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_85 (BatchNo (None, 9, 9, 128)    384         conv2d_94[0][0]                  
__________________________________________________________________________________________________
activation_85 (Activation)      (None, 9, 9, 128)    0           batch_normalization_85[0][0]     
__________________________________________________________________________________________________
conv2d_95 (Conv2D)              (None, 9, 9, 160)    143360      activation_85[0][0]              
__________________________________________________________________________________________________
batch_normalization_86 (BatchNo (None, 9, 9, 160)    480         conv2d_95[0][0]                  
__________________________________________________________________________________________________
activation_86 (Activation)      (None, 9, 9, 160)    0           batch_normalization_86[0][0]     
__________________________________________________________________________________________________
conv2d_93 (Conv2D)              (None, 9, 9, 192)    208896      block17_2_ac[0][0]               
__________________________________________________________________________________________________
conv2d_96 (Conv2D)              (None, 9, 9, 192)    215040      activation_86[0][0]              
__________________________________________________________________________________________________
batch_normalization_84 (BatchNo (None, 9, 9, 192)    576         conv2d_93[0][0]                  
__________________________________________________________________________________________________
batch_normalization_87 (BatchNo (None, 9, 9, 192)    576         conv2d_96[0][0]                  
__________________________________________________________________________________________________
activation_84 (Activation)      (None, 9, 9, 192)    0           batch_normalization_84[0][0]     
__________________________________________________________________________________________________
activation_87 (Activation)      (None, 9, 9, 192)    0           batch_normalization_87[0][0]     
__________________________________________________________________________________________________
block17_3_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_84[0][0]              
                                                                 activation_87[0][0]              
__________________________________________________________________________________________________
block17_3_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_3_mixed[0][0]            
__________________________________________________________________________________________________
block17_3 (Lambda)              (None, 9, 9, 1088)   0           block17_2_ac[0][0]               
                                                                 block17_3_conv[0][0]             
__________________________________________________________________________________________________
block17_3_ac (Activation)       (None, 9, 9, 1088)   0           block17_3[0][0]                  
__________________________________________________________________________________________________
conv2d_98 (Conv2D)              (None, 9, 9, 128)    139264      block17_3_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_89 (BatchNo (None, 9, 9, 128)    384         conv2d_98[0][0]                  
__________________________________________________________________________________________________
activation_89 (Activation)      (None, 9, 9, 128)    0           batch_normalization_89[0][0]     
__________________________________________________________________________________________________
conv2d_99 (Conv2D)              (None, 9, 9, 160)    143360      activation_89[0][0]              
__________________________________________________________________________________________________
batch_normalization_90 (BatchNo (None, 9, 9, 160)    480         conv2d_99[0][0]                  
__________________________________________________________________________________________________
activation_90 (Activation)      (None, 9, 9, 160)    0           batch_normalization_90[0][0]     
__________________________________________________________________________________________________
conv2d_97 (Conv2D)              (None, 9, 9, 192)    208896      block17_3_ac[0][0]               
__________________________________________________________________________________________________
conv2d_100 (Conv2D)             (None, 9, 9, 192)    215040      activation_90[0][0]              
__________________________________________________________________________________________________
batch_normalization_88 (BatchNo (None, 9, 9, 192)    576         conv2d_97[0][0]                  
__________________________________________________________________________________________________
batch_normalization_91 (BatchNo (None, 9, 9, 192)    576         conv2d_100[0][0]                 
__________________________________________________________________________________________________
activation_88 (Activation)      (None, 9, 9, 192)    0           batch_normalization_88[0][0]     
__________________________________________________________________________________________________
activation_91 (Activation)      (None, 9, 9, 192)    0           batch_normalization_91[0][0]     
__________________________________________________________________________________________________
block17_4_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_88[0][0]              
                                                                 activation_91[0][0]              
__________________________________________________________________________________________________
block17_4_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_4_mixed[0][0]            
__________________________________________________________________________________________________
block17_4 (Lambda)              (None, 9, 9, 1088)   0           block17_3_ac[0][0]               
                                                                 block17_4_conv[0][0]             
__________________________________________________________________________________________________
block17_4_ac (Activation)       (None, 9, 9, 1088)   0           block17_4[0][0]                  
__________________________________________________________________________________________________
conv2d_102 (Conv2D)             (None, 9, 9, 128)    139264      block17_4_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_93 (BatchNo (None, 9, 9, 128)    384         conv2d_102[0][0]                 
__________________________________________________________________________________________________
activation_93 (Activation)      (None, 9, 9, 128)    0           batch_normalization_93[0][0]     
__________________________________________________________________________________________________
conv2d_103 (Conv2D)             (None, 9, 9, 160)    143360      activation_93[0][0]              
__________________________________________________________________________________________________
batch_normalization_94 (BatchNo (None, 9, 9, 160)    480         conv2d_103[0][0]                 
__________________________________________________________________________________________________
activation_94 (Activation)      (None, 9, 9, 160)    0           batch_normalization_94[0][0]     
__________________________________________________________________________________________________
conv2d_101 (Conv2D)             (None, 9, 9, 192)    208896      block17_4_ac[0][0]               
__________________________________________________________________________________________________
conv2d_104 (Conv2D)             (None, 9, 9, 192)    215040      activation_94[0][0]              
__________________________________________________________________________________________________
batch_normalization_92 (BatchNo (None, 9, 9, 192)    576         conv2d_101[0][0]                 
__________________________________________________________________________________________________
batch_normalization_95 (BatchNo (None, 9, 9, 192)    576         conv2d_104[0][0]                 
__________________________________________________________________________________________________
activation_92 (Activation)      (None, 9, 9, 192)    0           batch_normalization_92[0][0]     
__________________________________________________________________________________________________
activation_95 (Activation)      (None, 9, 9, 192)    0           batch_normalization_95[0][0]     
__________________________________________________________________________________________________
block17_5_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_92[0][0]              
                                                                 activation_95[0][0]              
__________________________________________________________________________________________________
block17_5_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_5_mixed[0][0]            
__________________________________________________________________________________________________
block17_5 (Lambda)              (None, 9, 9, 1088)   0           block17_4_ac[0][0]               
                                                                 block17_5_conv[0][0]             
__________________________________________________________________________________________________
block17_5_ac (Activation)       (None, 9, 9, 1088)   0           block17_5[0][0]                  
__________________________________________________________________________________________________
conv2d_106 (Conv2D)             (None, 9, 9, 128)    139264      block17_5_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_97 (BatchNo (None, 9, 9, 128)    384         conv2d_106[0][0]                 
__________________________________________________________________________________________________
activation_97 (Activation)      (None, 9, 9, 128)    0           batch_normalization_97[0][0]     
__________________________________________________________________________________________________
conv2d_107 (Conv2D)             (None, 9, 9, 160)    143360      activation_97[0][0]              
__________________________________________________________________________________________________
batch_normalization_98 (BatchNo (None, 9, 9, 160)    480         conv2d_107[0][0]                 
__________________________________________________________________________________________________
activation_98 (Activation)      (None, 9, 9, 160)    0           batch_normalization_98[0][0]     
__________________________________________________________________________________________________
conv2d_105 (Conv2D)             (None, 9, 9, 192)    208896      block17_5_ac[0][0]               
__________________________________________________________________________________________________
conv2d_108 (Conv2D)             (None, 9, 9, 192)    215040      activation_98[0][0]              
__________________________________________________________________________________________________
batch_normalization_96 (BatchNo (None, 9, 9, 192)    576         conv2d_105[0][0]                 
__________________________________________________________________________________________________
batch_normalization_99 (BatchNo (None, 9, 9, 192)    576         conv2d_108[0][0]                 
__________________________________________________________________________________________________
activation_96 (Activation)      (None, 9, 9, 192)    0           batch_normalization_96[0][0]     
__________________________________________________________________________________________________
activation_99 (Activation)      (None, 9, 9, 192)    0           batch_normalization_99[0][0]     
__________________________________________________________________________________________________
block17_6_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_96[0][0]              
                                                                 activation_99[0][0]              
__________________________________________________________________________________________________
block17_6_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_6_mixed[0][0]            
__________________________________________________________________________________________________
block17_6 (Lambda)              (None, 9, 9, 1088)   0           block17_5_ac[0][0]               
                                                                 block17_6_conv[0][0]             
__________________________________________________________________________________________________
block17_6_ac (Activation)       (None, 9, 9, 1088)   0           block17_6[0][0]                  
__________________________________________________________________________________________________
conv2d_110 (Conv2D)             (None, 9, 9, 128)    139264      block17_6_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_101 (BatchN (None, 9, 9, 128)    384         conv2d_110[0][0]                 
__________________________________________________________________________________________________
activation_101 (Activation)     (None, 9, 9, 128)    0           batch_normalization_101[0][0]    
__________________________________________________________________________________________________
conv2d_111 (Conv2D)             (None, 9, 9, 160)    143360      activation_101[0][0]             
__________________________________________________________________________________________________
batch_normalization_102 (BatchN (None, 9, 9, 160)    480         conv2d_111[0][0]                 
__________________________________________________________________________________________________
activation_102 (Activation)     (None, 9, 9, 160)    0           batch_normalization_102[0][0]    
__________________________________________________________________________________________________
conv2d_109 (Conv2D)             (None, 9, 9, 192)    208896      block17_6_ac[0][0]               
__________________________________________________________________________________________________
conv2d_112 (Conv2D)             (None, 9, 9, 192)    215040      activation_102[0][0]             
__________________________________________________________________________________________________
batch_normalization_100 (BatchN (None, 9, 9, 192)    576         conv2d_109[0][0]                 
__________________________________________________________________________________________________
batch_normalization_103 (BatchN (None, 9, 9, 192)    576         conv2d_112[0][0]                 
__________________________________________________________________________________________________
activation_100 (Activation)     (None, 9, 9, 192)    0           batch_normalization_100[0][0]    
__________________________________________________________________________________________________
activation_103 (Activation)     (None, 9, 9, 192)    0           batch_normalization_103[0][0]    
__________________________________________________________________________________________________
block17_7_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_100[0][0]             
                                                                 activation_103[0][0]             
__________________________________________________________________________________________________
block17_7_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_7_mixed[0][0]            
__________________________________________________________________________________________________
block17_7 (Lambda)              (None, 9, 9, 1088)   0           block17_6_ac[0][0]               
                                                                 block17_7_conv[0][0]             
__________________________________________________________________________________________________
block17_7_ac (Activation)       (None, 9, 9, 1088)   0           block17_7[0][0]                  
__________________________________________________________________________________________________
conv2d_114 (Conv2D)             (None, 9, 9, 128)    139264      block17_7_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_105 (BatchN (None, 9, 9, 128)    384         conv2d_114[0][0]                 
__________________________________________________________________________________________________
activation_105 (Activation)     (None, 9, 9, 128)    0           batch_normalization_105[0][0]    
__________________________________________________________________________________________________
conv2d_115 (Conv2D)             (None, 9, 9, 160)    143360      activation_105[0][0]             
__________________________________________________________________________________________________
batch_normalization_106 (BatchN (None, 9, 9, 160)    480         conv2d_115[0][0]                 
__________________________________________________________________________________________________
activation_106 (Activation)     (None, 9, 9, 160)    0           batch_normalization_106[0][0]    
__________________________________________________________________________________________________
conv2d_113 (Conv2D)             (None, 9, 9, 192)    208896      block17_7_ac[0][0]               
__________________________________________________________________________________________________
conv2d_116 (Conv2D)             (None, 9, 9, 192)    215040      activation_106[0][0]             
__________________________________________________________________________________________________
batch_normalization_104 (BatchN (None, 9, 9, 192)    576         conv2d_113[0][0]                 
__________________________________________________________________________________________________
batch_normalization_107 (BatchN (None, 9, 9, 192)    576         conv2d_116[0][0]                 
__________________________________________________________________________________________________
activation_104 (Activation)     (None, 9, 9, 192)    0           batch_normalization_104[0][0]    
__________________________________________________________________________________________________
activation_107 (Activation)     (None, 9, 9, 192)    0           batch_normalization_107[0][0]    
__________________________________________________________________________________________________
block17_8_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_104[0][0]             
                                                                 activation_107[0][0]             
__________________________________________________________________________________________________
block17_8_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_8_mixed[0][0]            
__________________________________________________________________________________________________
block17_8 (Lambda)              (None, 9, 9, 1088)   0           block17_7_ac[0][0]               
                                                                 block17_8_conv[0][0]             
__________________________________________________________________________________________________
block17_8_ac (Activation)       (None, 9, 9, 1088)   0           block17_8[0][0]                  
__________________________________________________________________________________________________
conv2d_118 (Conv2D)             (None, 9, 9, 128)    139264      block17_8_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_109 (BatchN (None, 9, 9, 128)    384         conv2d_118[0][0]                 
__________________________________________________________________________________________________
activation_109 (Activation)     (None, 9, 9, 128)    0           batch_normalization_109[0][0]    
__________________________________________________________________________________________________
conv2d_119 (Conv2D)             (None, 9, 9, 160)    143360      activation_109[0][0]             
__________________________________________________________________________________________________
batch_normalization_110 (BatchN (None, 9, 9, 160)    480         conv2d_119[0][0]                 
__________________________________________________________________________________________________
activation_110 (Activation)     (None, 9, 9, 160)    0           batch_normalization_110[0][0]    
__________________________________________________________________________________________________
conv2d_117 (Conv2D)             (None, 9, 9, 192)    208896      block17_8_ac[0][0]               
__________________________________________________________________________________________________
conv2d_120 (Conv2D)             (None, 9, 9, 192)    215040      activation_110[0][0]             
__________________________________________________________________________________________________
batch_normalization_108 (BatchN (None, 9, 9, 192)    576         conv2d_117[0][0]                 
__________________________________________________________________________________________________
batch_normalization_111 (BatchN (None, 9, 9, 192)    576         conv2d_120[0][0]                 
__________________________________________________________________________________________________
activation_108 (Activation)     (None, 9, 9, 192)    0           batch_normalization_108[0][0]    
__________________________________________________________________________________________________
activation_111 (Activation)     (None, 9, 9, 192)    0           batch_normalization_111[0][0]    
__________________________________________________________________________________________________
block17_9_mixed (Concatenate)   (None, 9, 9, 384)    0           activation_108[0][0]             
                                                                 activation_111[0][0]             
__________________________________________________________________________________________________
block17_9_conv (Conv2D)         (None, 9, 9, 1088)   418880      block17_9_mixed[0][0]            
__________________________________________________________________________________________________
block17_9 (Lambda)              (None, 9, 9, 1088)   0           block17_8_ac[0][0]               
                                                                 block17_9_conv[0][0]             
__________________________________________________________________________________________________
block17_9_ac (Activation)       (None, 9, 9, 1088)   0           block17_9[0][0]                  
__________________________________________________________________________________________________
conv2d_122 (Conv2D)             (None, 9, 9, 128)    139264      block17_9_ac[0][0]               
__________________________________________________________________________________________________
batch_normalization_113 (BatchN (None, 9, 9, 128)    384         conv2d_122[0][0]                 
__________________________________________________________________________________________________
activation_113 (Activation)     (None, 9, 9, 128)    0           batch_normalization_113[0][0]    
__________________________________________________________________________________________________
conv2d_123 (Conv2D)             (None, 9, 9, 160)    143360      activation_113[0][0]             
__________________________________________________________________________________________________
batch_normalization_114 (BatchN (None, 9, 9, 160)    480         conv2d_123[0][0]                 
__________________________________________________________________________________________________
activation_114 (Activation)     (None, 9, 9, 160)    0           batch_normalization_114[0][0]    
__________________________________________________________________________________________________
conv2d_121 (Conv2D)             (None, 9, 9, 192)    208896      block17_9_ac[0][0]               
__________________________________________________________________________________________________
conv2d_124 (Conv2D)             (None, 9, 9, 192)    215040      activation_114[0][0]             
__________________________________________________________________________________________________
batch_normalization_112 (BatchN (None, 9, 9, 192)    576         conv2d_121[0][0]                 
__________________________________________________________________________________________________
batch_normalization_115 (BatchN (None, 9, 9, 192)    576         conv2d_124[0][0]                 
__________________________________________________________________________________________________
activation_112 (Activation)     (None, 9, 9, 192)    0           batch_normalization_112[0][0]    
__________________________________________________________________________________________________
activation_115 (Activation)     (None, 9, 9, 192)    0           batch_normalization_115[0][0]    
__________________________________________________________________________________________________
block17_10_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_112[0][0]             
                                                                 activation_115[0][0]             
__________________________________________________________________________________________________
block17_10_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_10_mixed[0][0]           
__________________________________________________________________________________________________
block17_10 (Lambda)             (None, 9, 9, 1088)   0           block17_9_ac[0][0]               
                                                                 block17_10_conv[0][0]            
__________________________________________________________________________________________________
block17_10_ac (Activation)      (None, 9, 9, 1088)   0           block17_10[0][0]                 
__________________________________________________________________________________________________
conv2d_126 (Conv2D)             (None, 9, 9, 128)    139264      block17_10_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_117 (BatchN (None, 9, 9, 128)    384         conv2d_126[0][0]                 
__________________________________________________________________________________________________
activation_117 (Activation)     (None, 9, 9, 128)    0           batch_normalization_117[0][0]    
__________________________________________________________________________________________________
conv2d_127 (Conv2D)             (None, 9, 9, 160)    143360      activation_117[0][0]             
__________________________________________________________________________________________________
batch_normalization_118 (BatchN (None, 9, 9, 160)    480         conv2d_127[0][0]                 
__________________________________________________________________________________________________
activation_118 (Activation)     (None, 9, 9, 160)    0           batch_normalization_118[0][0]    
__________________________________________________________________________________________________
conv2d_125 (Conv2D)             (None, 9, 9, 192)    208896      block17_10_ac[0][0]              
__________________________________________________________________________________________________
conv2d_128 (Conv2D)             (None, 9, 9, 192)    215040      activation_118[0][0]             
__________________________________________________________________________________________________
batch_normalization_116 (BatchN (None, 9, 9, 192)    576         conv2d_125[0][0]                 
__________________________________________________________________________________________________
batch_normalization_119 (BatchN (None, 9, 9, 192)    576         conv2d_128[0][0]                 
__________________________________________________________________________________________________
activation_116 (Activation)     (None, 9, 9, 192)    0           batch_normalization_116[0][0]    
__________________________________________________________________________________________________
activation_119 (Activation)     (None, 9, 9, 192)    0           batch_normalization_119[0][0]    
__________________________________________________________________________________________________
block17_11_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_116[0][0]             
                                                                 activation_119[0][0]             
__________________________________________________________________________________________________
block17_11_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_11_mixed[0][0]           
__________________________________________________________________________________________________
block17_11 (Lambda)             (None, 9, 9, 1088)   0           block17_10_ac[0][0]              
                                                                 block17_11_conv[0][0]            
__________________________________________________________________________________________________
block17_11_ac (Activation)      (None, 9, 9, 1088)   0           block17_11[0][0]                 
__________________________________________________________________________________________________
conv2d_130 (Conv2D)             (None, 9, 9, 128)    139264      block17_11_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_121 (BatchN (None, 9, 9, 128)    384         conv2d_130[0][0]                 
__________________________________________________________________________________________________
activation_121 (Activation)     (None, 9, 9, 128)    0           batch_normalization_121[0][0]    
__________________________________________________________________________________________________
conv2d_131 (Conv2D)             (None, 9, 9, 160)    143360      activation_121[0][0]             
__________________________________________________________________________________________________
batch_normalization_122 (BatchN (None, 9, 9, 160)    480         conv2d_131[0][0]                 
__________________________________________________________________________________________________
activation_122 (Activation)     (None, 9, 9, 160)    0           batch_normalization_122[0][0]    
__________________________________________________________________________________________________
conv2d_129 (Conv2D)             (None, 9, 9, 192)    208896      block17_11_ac[0][0]              
__________________________________________________________________________________________________
conv2d_132 (Conv2D)             (None, 9, 9, 192)    215040      activation_122[0][0]             
__________________________________________________________________________________________________
batch_normalization_120 (BatchN (None, 9, 9, 192)    576         conv2d_129[0][0]                 
__________________________________________________________________________________________________
batch_normalization_123 (BatchN (None, 9, 9, 192)    576         conv2d_132[0][0]                 
__________________________________________________________________________________________________
activation_120 (Activation)     (None, 9, 9, 192)    0           batch_normalization_120[0][0]    
__________________________________________________________________________________________________
activation_123 (Activation)     (None, 9, 9, 192)    0           batch_normalization_123[0][0]    
__________________________________________________________________________________________________
block17_12_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_120[0][0]             
                                                                 activation_123[0][0]             
__________________________________________________________________________________________________
block17_12_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_12_mixed[0][0]           
__________________________________________________________________________________________________
block17_12 (Lambda)             (None, 9, 9, 1088)   0           block17_11_ac[0][0]              
                                                                 block17_12_conv[0][0]            
__________________________________________________________________________________________________
block17_12_ac (Activation)      (None, 9, 9, 1088)   0           block17_12[0][0]                 
__________________________________________________________________________________________________
conv2d_134 (Conv2D)             (None, 9, 9, 128)    139264      block17_12_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_125 (BatchN (None, 9, 9, 128)    384         conv2d_134[0][0]                 
__________________________________________________________________________________________________
activation_125 (Activation)     (None, 9, 9, 128)    0           batch_normalization_125[0][0]    
__________________________________________________________________________________________________
conv2d_135 (Conv2D)             (None, 9, 9, 160)    143360      activation_125[0][0]             
__________________________________________________________________________________________________
batch_normalization_126 (BatchN (None, 9, 9, 160)    480         conv2d_135[0][0]                 
__________________________________________________________________________________________________
activation_126 (Activation)     (None, 9, 9, 160)    0           batch_normalization_126[0][0]    
__________________________________________________________________________________________________
conv2d_133 (Conv2D)             (None, 9, 9, 192)    208896      block17_12_ac[0][0]              
__________________________________________________________________________________________________
conv2d_136 (Conv2D)             (None, 9, 9, 192)    215040      activation_126[0][0]             
__________________________________________________________________________________________________
batch_normalization_124 (BatchN (None, 9, 9, 192)    576         conv2d_133[0][0]                 
__________________________________________________________________________________________________
batch_normalization_127 (BatchN (None, 9, 9, 192)    576         conv2d_136[0][0]                 
__________________________________________________________________________________________________
activation_124 (Activation)     (None, 9, 9, 192)    0           batch_normalization_124[0][0]    
__________________________________________________________________________________________________
activation_127 (Activation)     (None, 9, 9, 192)    0           batch_normalization_127[0][0]    
__________________________________________________________________________________________________
block17_13_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_124[0][0]             
                                                                 activation_127[0][0]             
__________________________________________________________________________________________________
block17_13_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_13_mixed[0][0]           
__________________________________________________________________________________________________
block17_13 (Lambda)             (None, 9, 9, 1088)   0           block17_12_ac[0][0]              
                                                                 block17_13_conv[0][0]            
__________________________________________________________________________________________________
block17_13_ac (Activation)      (None, 9, 9, 1088)   0           block17_13[0][0]                 
__________________________________________________________________________________________________
conv2d_138 (Conv2D)             (None, 9, 9, 128)    139264      block17_13_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_129 (BatchN (None, 9, 9, 128)    384         conv2d_138[0][0]                 
__________________________________________________________________________________________________
activation_129 (Activation)     (None, 9, 9, 128)    0           batch_normalization_129[0][0]    
__________________________________________________________________________________________________
conv2d_139 (Conv2D)             (None, 9, 9, 160)    143360      activation_129[0][0]             
__________________________________________________________________________________________________
batch_normalization_130 (BatchN (None, 9, 9, 160)    480         conv2d_139[0][0]                 
__________________________________________________________________________________________________
activation_130 (Activation)     (None, 9, 9, 160)    0           batch_normalization_130[0][0]    
__________________________________________________________________________________________________
conv2d_137 (Conv2D)             (None, 9, 9, 192)    208896      block17_13_ac[0][0]              
__________________________________________________________________________________________________
conv2d_140 (Conv2D)             (None, 9, 9, 192)    215040      activation_130[0][0]             
__________________________________________________________________________________________________
batch_normalization_128 (BatchN (None, 9, 9, 192)    576         conv2d_137[0][0]                 
__________________________________________________________________________________________________
batch_normalization_131 (BatchN (None, 9, 9, 192)    576         conv2d_140[0][0]                 
__________________________________________________________________________________________________
activation_128 (Activation)     (None, 9, 9, 192)    0           batch_normalization_128[0][0]    
__________________________________________________________________________________________________
activation_131 (Activation)     (None, 9, 9, 192)    0           batch_normalization_131[0][0]    
__________________________________________________________________________________________________
block17_14_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_128[0][0]             
                                                                 activation_131[0][0]             
__________________________________________________________________________________________________
block17_14_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_14_mixed[0][0]           
__________________________________________________________________________________________________
block17_14 (Lambda)             (None, 9, 9, 1088)   0           block17_13_ac[0][0]              
                                                                 block17_14_conv[0][0]            
__________________________________________________________________________________________________
block17_14_ac (Activation)      (None, 9, 9, 1088)   0           block17_14[0][0]                 
__________________________________________________________________________________________________
conv2d_142 (Conv2D)             (None, 9, 9, 128)    139264      block17_14_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_133 (BatchN (None, 9, 9, 128)    384         conv2d_142[0][0]                 
__________________________________________________________________________________________________
activation_133 (Activation)     (None, 9, 9, 128)    0           batch_normalization_133[0][0]    
__________________________________________________________________________________________________
conv2d_143 (Conv2D)             (None, 9, 9, 160)    143360      activation_133[0][0]             
__________________________________________________________________________________________________
batch_normalization_134 (BatchN (None, 9, 9, 160)    480         conv2d_143[0][0]                 
__________________________________________________________________________________________________
activation_134 (Activation)     (None, 9, 9, 160)    0           batch_normalization_134[0][0]    
__________________________________________________________________________________________________
conv2d_141 (Conv2D)             (None, 9, 9, 192)    208896      block17_14_ac[0][0]              
__________________________________________________________________________________________________
conv2d_144 (Conv2D)             (None, 9, 9, 192)    215040      activation_134[0][0]             
__________________________________________________________________________________________________
batch_normalization_132 (BatchN (None, 9, 9, 192)    576         conv2d_141[0][0]                 
__________________________________________________________________________________________________
batch_normalization_135 (BatchN (None, 9, 9, 192)    576         conv2d_144[0][0]                 
__________________________________________________________________________________________________
activation_132 (Activation)     (None, 9, 9, 192)    0           batch_normalization_132[0][0]    
__________________________________________________________________________________________________
activation_135 (Activation)     (None, 9, 9, 192)    0           batch_normalization_135[0][0]    
__________________________________________________________________________________________________
block17_15_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_132[0][0]             
                                                                 activation_135[0][0]             
__________________________________________________________________________________________________
block17_15_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_15_mixed[0][0]           
__________________________________________________________________________________________________
block17_15 (Lambda)             (None, 9, 9, 1088)   0           block17_14_ac[0][0]              
                                                                 block17_15_conv[0][0]            
__________________________________________________________________________________________________
block17_15_ac (Activation)      (None, 9, 9, 1088)   0           block17_15[0][0]                 
__________________________________________________________________________________________________
conv2d_146 (Conv2D)             (None, 9, 9, 128)    139264      block17_15_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_137 (BatchN (None, 9, 9, 128)    384         conv2d_146[0][0]                 
__________________________________________________________________________________________________
activation_137 (Activation)     (None, 9, 9, 128)    0           batch_normalization_137[0][0]    
__________________________________________________________________________________________________
conv2d_147 (Conv2D)             (None, 9, 9, 160)    143360      activation_137[0][0]             
__________________________________________________________________________________________________
batch_normalization_138 (BatchN (None, 9, 9, 160)    480         conv2d_147[0][0]                 
__________________________________________________________________________________________________
activation_138 (Activation)     (None, 9, 9, 160)    0           batch_normalization_138[0][0]    
__________________________________________________________________________________________________
conv2d_145 (Conv2D)             (None, 9, 9, 192)    208896      block17_15_ac[0][0]              
__________________________________________________________________________________________________
conv2d_148 (Conv2D)             (None, 9, 9, 192)    215040      activation_138[0][0]             
__________________________________________________________________________________________________
batch_normalization_136 (BatchN (None, 9, 9, 192)    576         conv2d_145[0][0]                 
__________________________________________________________________________________________________
batch_normalization_139 (BatchN (None, 9, 9, 192)    576         conv2d_148[0][0]                 
__________________________________________________________________________________________________
activation_136 (Activation)     (None, 9, 9, 192)    0           batch_normalization_136[0][0]    
__________________________________________________________________________________________________
activation_139 (Activation)     (None, 9, 9, 192)    0           batch_normalization_139[0][0]    
__________________________________________________________________________________________________
block17_16_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_136[0][0]             
                                                                 activation_139[0][0]             
__________________________________________________________________________________________________
block17_16_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_16_mixed[0][0]           
__________________________________________________________________________________________________
block17_16 (Lambda)             (None, 9, 9, 1088)   0           block17_15_ac[0][0]              
                                                                 block17_16_conv[0][0]            
__________________________________________________________________________________________________
block17_16_ac (Activation)      (None, 9, 9, 1088)   0           block17_16[0][0]                 
__________________________________________________________________________________________________
conv2d_150 (Conv2D)             (None, 9, 9, 128)    139264      block17_16_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_141 (BatchN (None, 9, 9, 128)    384         conv2d_150[0][0]                 
__________________________________________________________________________________________________
activation_141 (Activation)     (None, 9, 9, 128)    0           batch_normalization_141[0][0]    
__________________________________________________________________________________________________
conv2d_151 (Conv2D)             (None, 9, 9, 160)    143360      activation_141[0][0]             
__________________________________________________________________________________________________
batch_normalization_142 (BatchN (None, 9, 9, 160)    480         conv2d_151[0][0]                 
__________________________________________________________________________________________________
activation_142 (Activation)     (None, 9, 9, 160)    0           batch_normalization_142[0][0]    
__________________________________________________________________________________________________
conv2d_149 (Conv2D)             (None, 9, 9, 192)    208896      block17_16_ac[0][0]              
__________________________________________________________________________________________________
conv2d_152 (Conv2D)             (None, 9, 9, 192)    215040      activation_142[0][0]             
__________________________________________________________________________________________________
batch_normalization_140 (BatchN (None, 9, 9, 192)    576         conv2d_149[0][0]                 
__________________________________________________________________________________________________
batch_normalization_143 (BatchN (None, 9, 9, 192)    576         conv2d_152[0][0]                 
__________________________________________________________________________________________________
activation_140 (Activation)     (None, 9, 9, 192)    0           batch_normalization_140[0][0]    
__________________________________________________________________________________________________
activation_143 (Activation)     (None, 9, 9, 192)    0           batch_normalization_143[0][0]    
__________________________________________________________________________________________________
block17_17_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_140[0][0]             
                                                                 activation_143[0][0]             
__________________________________________________________________________________________________
block17_17_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_17_mixed[0][0]           
__________________________________________________________________________________________________
block17_17 (Lambda)             (None, 9, 9, 1088)   0           block17_16_ac[0][0]              
                                                                 block17_17_conv[0][0]            
__________________________________________________________________________________________________
block17_17_ac (Activation)      (None, 9, 9, 1088)   0           block17_17[0][0]                 
__________________________________________________________________________________________________
conv2d_154 (Conv2D)             (None, 9, 9, 128)    139264      block17_17_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_145 (BatchN (None, 9, 9, 128)    384         conv2d_154[0][0]                 
__________________________________________________________________________________________________
activation_145 (Activation)     (None, 9, 9, 128)    0           batch_normalization_145[0][0]    
__________________________________________________________________________________________________
conv2d_155 (Conv2D)             (None, 9, 9, 160)    143360      activation_145[0][0]             
__________________________________________________________________________________________________
batch_normalization_146 (BatchN (None, 9, 9, 160)    480         conv2d_155[0][0]                 
__________________________________________________________________________________________________
activation_146 (Activation)     (None, 9, 9, 160)    0           batch_normalization_146[0][0]    
__________________________________________________________________________________________________
conv2d_153 (Conv2D)             (None, 9, 9, 192)    208896      block17_17_ac[0][0]              
__________________________________________________________________________________________________
conv2d_156 (Conv2D)             (None, 9, 9, 192)    215040      activation_146[0][0]             
__________________________________________________________________________________________________
batch_normalization_144 (BatchN (None, 9, 9, 192)    576         conv2d_153[0][0]                 
__________________________________________________________________________________________________
batch_normalization_147 (BatchN (None, 9, 9, 192)    576         conv2d_156[0][0]                 
__________________________________________________________________________________________________
activation_144 (Activation)     (None, 9, 9, 192)    0           batch_normalization_144[0][0]    
__________________________________________________________________________________________________
activation_147 (Activation)     (None, 9, 9, 192)    0           batch_normalization_147[0][0]    
__________________________________________________________________________________________________
block17_18_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_144[0][0]             
                                                                 activation_147[0][0]             
__________________________________________________________________________________________________
block17_18_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_18_mixed[0][0]           
__________________________________________________________________________________________________
block17_18 (Lambda)             (None, 9, 9, 1088)   0           block17_17_ac[0][0]              
                                                                 block17_18_conv[0][0]            
__________________________________________________________________________________________________
block17_18_ac (Activation)      (None, 9, 9, 1088)   0           block17_18[0][0]                 
__________________________________________________________________________________________________
conv2d_158 (Conv2D)             (None, 9, 9, 128)    139264      block17_18_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_149 (BatchN (None, 9, 9, 128)    384         conv2d_158[0][0]                 
__________________________________________________________________________________________________
activation_149 (Activation)     (None, 9, 9, 128)    0           batch_normalization_149[0][0]    
__________________________________________________________________________________________________
conv2d_159 (Conv2D)             (None, 9, 9, 160)    143360      activation_149[0][0]             
__________________________________________________________________________________________________
batch_normalization_150 (BatchN (None, 9, 9, 160)    480         conv2d_159[0][0]                 
__________________________________________________________________________________________________
activation_150 (Activation)     (None, 9, 9, 160)    0           batch_normalization_150[0][0]    
__________________________________________________________________________________________________
conv2d_157 (Conv2D)             (None, 9, 9, 192)    208896      block17_18_ac[0][0]              
__________________________________________________________________________________________________
conv2d_160 (Conv2D)             (None, 9, 9, 192)    215040      activation_150[0][0]             
__________________________________________________________________________________________________
batch_normalization_148 (BatchN (None, 9, 9, 192)    576         conv2d_157[0][0]                 
__________________________________________________________________________________________________
batch_normalization_151 (BatchN (None, 9, 9, 192)    576         conv2d_160[0][0]                 
__________________________________________________________________________________________________
activation_148 (Activation)     (None, 9, 9, 192)    0           batch_normalization_148[0][0]    
__________________________________________________________________________________________________
activation_151 (Activation)     (None, 9, 9, 192)    0           batch_normalization_151[0][0]    
__________________________________________________________________________________________________
block17_19_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_148[0][0]             
                                                                 activation_151[0][0]             
__________________________________________________________________________________________________
block17_19_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_19_mixed[0][0]           
__________________________________________________________________________________________________
block17_19 (Lambda)             (None, 9, 9, 1088)   0           block17_18_ac[0][0]              
                                                                 block17_19_conv[0][0]            
__________________________________________________________________________________________________
block17_19_ac (Activation)      (None, 9, 9, 1088)   0           block17_19[0][0]                 
__________________________________________________________________________________________________
conv2d_162 (Conv2D)             (None, 9, 9, 128)    139264      block17_19_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_153 (BatchN (None, 9, 9, 128)    384         conv2d_162[0][0]                 
__________________________________________________________________________________________________
activation_153 (Activation)     (None, 9, 9, 128)    0           batch_normalization_153[0][0]    
__________________________________________________________________________________________________
conv2d_163 (Conv2D)             (None, 9, 9, 160)    143360      activation_153[0][0]             
__________________________________________________________________________________________________
batch_normalization_154 (BatchN (None, 9, 9, 160)    480         conv2d_163[0][0]                 
__________________________________________________________________________________________________
activation_154 (Activation)     (None, 9, 9, 160)    0           batch_normalization_154[0][0]    
__________________________________________________________________________________________________
conv2d_161 (Conv2D)             (None, 9, 9, 192)    208896      block17_19_ac[0][0]              
__________________________________________________________________________________________________
conv2d_164 (Conv2D)             (None, 9, 9, 192)    215040      activation_154[0][0]             
__________________________________________________________________________________________________
batch_normalization_152 (BatchN (None, 9, 9, 192)    576         conv2d_161[0][0]                 
__________________________________________________________________________________________________
batch_normalization_155 (BatchN (None, 9, 9, 192)    576         conv2d_164[0][0]                 
__________________________________________________________________________________________________
activation_152 (Activation)     (None, 9, 9, 192)    0           batch_normalization_152[0][0]    
__________________________________________________________________________________________________
activation_155 (Activation)     (None, 9, 9, 192)    0           batch_normalization_155[0][0]    
__________________________________________________________________________________________________
block17_20_mixed (Concatenate)  (None, 9, 9, 384)    0           activation_152[0][0]             
                                                                 activation_155[0][0]             
__________________________________________________________________________________________________
block17_20_conv (Conv2D)        (None, 9, 9, 1088)   418880      block17_20_mixed[0][0]           
__________________________________________________________________________________________________
block17_20 (Lambda)             (None, 9, 9, 1088)   0           block17_19_ac[0][0]              
                                                                 block17_20_conv[0][0]            
__________________________________________________________________________________________________
block17_20_ac (Activation)      (None, 9, 9, 1088)   0           block17_20[0][0]                 
__________________________________________________________________________________________________
conv2d_169 (Conv2D)             (None, 9, 9, 256)    278528      block17_20_ac[0][0]              
__________________________________________________________________________________________________
batch_normalization_160 (BatchN (None, 9, 9, 256)    768         conv2d_169[0][0]                 
__________________________________________________________________________________________________
activation_160 (Activation)     (None, 9, 9, 256)    0           batch_normalization_160[0][0]    
__________________________________________________________________________________________________
conv2d_165 (Conv2D)             (None, 9, 9, 256)    278528      block17_20_ac[0][0]              
__________________________________________________________________________________________________
conv2d_167 (Conv2D)             (None, 9, 9, 256)    278528      block17_20_ac[0][0]              
__________________________________________________________________________________________________
conv2d_170 (Conv2D)             (None, 9, 9, 288)    663552      activation_160[0][0]             
__________________________________________________________________________________________________
batch_normalization_156 (BatchN (None, 9, 9, 256)    768         conv2d_165[0][0]                 
__________________________________________________________________________________________________
batch_normalization_158 (BatchN (None, 9, 9, 256)    768         conv2d_167[0][0]                 
__________________________________________________________________________________________________
batch_normalization_161 (BatchN (None, 9, 9, 288)    864         conv2d_170[0][0]                 
__________________________________________________________________________________________________
activation_156 (Activation)     (None, 9, 9, 256)    0           batch_normalization_156[0][0]    
__________________________________________________________________________________________________
activation_158 (Activation)     (None, 9, 9, 256)    0           batch_normalization_158[0][0]    
__________________________________________________________________________________________________
activation_161 (Activation)     (None, 9, 9, 288)    0           batch_normalization_161[0][0]    
__________________________________________________________________________________________________
conv2d_166 (Conv2D)             (None, 4, 4, 384)    884736      activation_156[0][0]             
__________________________________________________________________________________________________
conv2d_168 (Conv2D)             (None, 4, 4, 288)    663552      activation_158[0][0]             
__________________________________________________________________________________________________
conv2d_171 (Conv2D)             (None, 4, 4, 320)    829440      activation_161[0][0]             
__________________________________________________________________________________________________
batch_normalization_157 (BatchN (None, 4, 4, 384)    1152        conv2d_166[0][0]                 
__________________________________________________________________________________________________
batch_normalization_159 (BatchN (None, 4, 4, 288)    864         conv2d_168[0][0]                 
__________________________________________________________________________________________________
batch_normalization_162 (BatchN (None, 4, 4, 320)    960         conv2d_171[0][0]                 
__________________________________________________________________________________________________
activation_157 (Activation)     (None, 4, 4, 384)    0           batch_normalization_157[0][0]    
__________________________________________________________________________________________________
activation_159 (Activation)     (None, 4, 4, 288)    0           batch_normalization_159[0][0]    
__________________________________________________________________________________________________
activation_162 (Activation)     (None, 4, 4, 320)    0           batch_normalization_162[0][0]    
__________________________________________________________________________________________________
max_pooling2d_12 (MaxPooling2D) (None, 4, 4, 1088)   0           block17_20_ac[0][0]              
__________________________________________________________________________________________________
mixed_7a (Concatenate)          (None, 4, 4, 2080)   0           activation_157[0][0]             
                                                                 activation_159[0][0]             
                                                                 activation_162[0][0]             
                                                                 max_pooling2d_12[0][0]           
__________________________________________________________________________________________________
conv2d_173 (Conv2D)             (None, 4, 4, 192)    399360      mixed_7a[0][0]                   
__________________________________________________________________________________________________
batch_normalization_164 (BatchN (None, 4, 4, 192)    576         conv2d_173[0][0]                 
__________________________________________________________________________________________________
activation_164 (Activation)     (None, 4, 4, 192)    0           batch_normalization_164[0][0]    
__________________________________________________________________________________________________
conv2d_174 (Conv2D)             (None, 4, 4, 224)    129024      activation_164[0][0]             
__________________________________________________________________________________________________
batch_normalization_165 (BatchN (None, 4, 4, 224)    672         conv2d_174[0][0]                 
__________________________________________________________________________________________________
activation_165 (Activation)     (None, 4, 4, 224)    0           batch_normalization_165[0][0]    
__________________________________________________________________________________________________
conv2d_172 (Conv2D)             (None, 4, 4, 192)    399360      mixed_7a[0][0]                   
__________________________________________________________________________________________________
conv2d_175 (Conv2D)             (None, 4, 4, 256)    172032      activation_165[0][0]             
__________________________________________________________________________________________________
batch_normalization_163 (BatchN (None, 4, 4, 192)    576         conv2d_172[0][0]                 
__________________________________________________________________________________________________
batch_normalization_166 (BatchN (None, 4, 4, 256)    768         conv2d_175[0][0]                 
__________________________________________________________________________________________________
activation_163 (Activation)     (None, 4, 4, 192)    0           batch_normalization_163[0][0]    
__________________________________________________________________________________________________
activation_166 (Activation)     (None, 4, 4, 256)    0           batch_normalization_166[0][0]    
__________________________________________________________________________________________________
block8_1_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_163[0][0]             
                                                                 activation_166[0][0]             
__________________________________________________________________________________________________
block8_1_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_1_mixed[0][0]             
__________________________________________________________________________________________________
block8_1 (Lambda)               (None, 4, 4, 2080)   0           mixed_7a[0][0]                   
                                                                 block8_1_conv[0][0]              
__________________________________________________________________________________________________
block8_1_ac (Activation)        (None, 4, 4, 2080)   0           block8_1[0][0]                   
__________________________________________________________________________________________________
conv2d_177 (Conv2D)             (None, 4, 4, 192)    399360      block8_1_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_168 (BatchN (None, 4, 4, 192)    576         conv2d_177[0][0]                 
__________________________________________________________________________________________________
activation_168 (Activation)     (None, 4, 4, 192)    0           batch_normalization_168[0][0]    
__________________________________________________________________________________________________
conv2d_178 (Conv2D)             (None, 4, 4, 224)    129024      activation_168[0][0]             
__________________________________________________________________________________________________
batch_normalization_169 (BatchN (None, 4, 4, 224)    672         conv2d_178[0][0]                 
__________________________________________________________________________________________________
activation_169 (Activation)     (None, 4, 4, 224)    0           batch_normalization_169[0][0]    
__________________________________________________________________________________________________
conv2d_176 (Conv2D)             (None, 4, 4, 192)    399360      block8_1_ac[0][0]                
__________________________________________________________________________________________________
conv2d_179 (Conv2D)             (None, 4, 4, 256)    172032      activation_169[0][0]             
__________________________________________________________________________________________________
batch_normalization_167 (BatchN (None, 4, 4, 192)    576         conv2d_176[0][0]                 
__________________________________________________________________________________________________
batch_normalization_170 (BatchN (None, 4, 4, 256)    768         conv2d_179[0][0]                 
__________________________________________________________________________________________________
activation_167 (Activation)     (None, 4, 4, 192)    0           batch_normalization_167[0][0]    
__________________________________________________________________________________________________
activation_170 (Activation)     (None, 4, 4, 256)    0           batch_normalization_170[0][0]    
__________________________________________________________________________________________________
block8_2_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_167[0][0]             
                                                                 activation_170[0][0]             
__________________________________________________________________________________________________
block8_2_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_2_mixed[0][0]             
__________________________________________________________________________________________________
block8_2 (Lambda)               (None, 4, 4, 2080)   0           block8_1_ac[0][0]                
                                                                 block8_2_conv[0][0]              
__________________________________________________________________________________________________
block8_2_ac (Activation)        (None, 4, 4, 2080)   0           block8_2[0][0]                   
__________________________________________________________________________________________________
conv2d_181 (Conv2D)             (None, 4, 4, 192)    399360      block8_2_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_172 (BatchN (None, 4, 4, 192)    576         conv2d_181[0][0]                 
__________________________________________________________________________________________________
activation_172 (Activation)     (None, 4, 4, 192)    0           batch_normalization_172[0][0]    
__________________________________________________________________________________________________
conv2d_182 (Conv2D)             (None, 4, 4, 224)    129024      activation_172[0][0]             
__________________________________________________________________________________________________
batch_normalization_173 (BatchN (None, 4, 4, 224)    672         conv2d_182[0][0]                 
__________________________________________________________________________________________________
activation_173 (Activation)     (None, 4, 4, 224)    0           batch_normalization_173[0][0]    
__________________________________________________________________________________________________
conv2d_180 (Conv2D)             (None, 4, 4, 192)    399360      block8_2_ac[0][0]                
__________________________________________________________________________________________________
conv2d_183 (Conv2D)             (None, 4, 4, 256)    172032      activation_173[0][0]             
__________________________________________________________________________________________________
batch_normalization_171 (BatchN (None, 4, 4, 192)    576         conv2d_180[0][0]                 
__________________________________________________________________________________________________
batch_normalization_174 (BatchN (None, 4, 4, 256)    768         conv2d_183[0][0]                 
__________________________________________________________________________________________________
activation_171 (Activation)     (None, 4, 4, 192)    0           batch_normalization_171[0][0]    
__________________________________________________________________________________________________
activation_174 (Activation)     (None, 4, 4, 256)    0           batch_normalization_174[0][0]    
__________________________________________________________________________________________________
block8_3_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_171[0][0]             
                                                                 activation_174[0][0]             
__________________________________________________________________________________________________
block8_3_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_3_mixed[0][0]             
__________________________________________________________________________________________________
block8_3 (Lambda)               (None, 4, 4, 2080)   0           block8_2_ac[0][0]                
                                                                 block8_3_conv[0][0]              
__________________________________________________________________________________________________
block8_3_ac (Activation)        (None, 4, 4, 2080)   0           block8_3[0][0]                   
__________________________________________________________________________________________________
conv2d_185 (Conv2D)             (None, 4, 4, 192)    399360      block8_3_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_176 (BatchN (None, 4, 4, 192)    576         conv2d_185[0][0]                 
__________________________________________________________________________________________________
activation_176 (Activation)     (None, 4, 4, 192)    0           batch_normalization_176[0][0]    
__________________________________________________________________________________________________
conv2d_186 (Conv2D)             (None, 4, 4, 224)    129024      activation_176[0][0]             
__________________________________________________________________________________________________
batch_normalization_177 (BatchN (None, 4, 4, 224)    672         conv2d_186[0][0]                 
__________________________________________________________________________________________________
activation_177 (Activation)     (None, 4, 4, 224)    0           batch_normalization_177[0][0]    
__________________________________________________________________________________________________
conv2d_184 (Conv2D)             (None, 4, 4, 192)    399360      block8_3_ac[0][0]                
__________________________________________________________________________________________________
conv2d_187 (Conv2D)             (None, 4, 4, 256)    172032      activation_177[0][0]             
__________________________________________________________________________________________________
batch_normalization_175 (BatchN (None, 4, 4, 192)    576         conv2d_184[0][0]                 
__________________________________________________________________________________________________
batch_normalization_178 (BatchN (None, 4, 4, 256)    768         conv2d_187[0][0]                 
__________________________________________________________________________________________________
activation_175 (Activation)     (None, 4, 4, 192)    0           batch_normalization_175[0][0]    
__________________________________________________________________________________________________
activation_178 (Activation)     (None, 4, 4, 256)    0           batch_normalization_178[0][0]    
__________________________________________________________________________________________________
block8_4_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_175[0][0]             
                                                                 activation_178[0][0]             
__________________________________________________________________________________________________
block8_4_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_4_mixed[0][0]             
__________________________________________________________________________________________________
block8_4 (Lambda)               (None, 4, 4, 2080)   0           block8_3_ac[0][0]                
                                                                 block8_4_conv[0][0]              
__________________________________________________________________________________________________
block8_4_ac (Activation)        (None, 4, 4, 2080)   0           block8_4[0][0]                   
__________________________________________________________________________________________________
conv2d_189 (Conv2D)             (None, 4, 4, 192)    399360      block8_4_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_180 (BatchN (None, 4, 4, 192)    576         conv2d_189[0][0]                 
__________________________________________________________________________________________________
activation_180 (Activation)     (None, 4, 4, 192)    0           batch_normalization_180[0][0]    
__________________________________________________________________________________________________
conv2d_190 (Conv2D)             (None, 4, 4, 224)    129024      activation_180[0][0]             
__________________________________________________________________________________________________
batch_normalization_181 (BatchN (None, 4, 4, 224)    672         conv2d_190[0][0]                 
__________________________________________________________________________________________________
activation_181 (Activation)     (None, 4, 4, 224)    0           batch_normalization_181[0][0]    
__________________________________________________________________________________________________
conv2d_188 (Conv2D)             (None, 4, 4, 192)    399360      block8_4_ac[0][0]                
__________________________________________________________________________________________________
conv2d_191 (Conv2D)             (None, 4, 4, 256)    172032      activation_181[0][0]             
__________________________________________________________________________________________________
batch_normalization_179 (BatchN (None, 4, 4, 192)    576         conv2d_188[0][0]                 
__________________________________________________________________________________________________
batch_normalization_182 (BatchN (None, 4, 4, 256)    768         conv2d_191[0][0]                 
__________________________________________________________________________________________________
activation_179 (Activation)     (None, 4, 4, 192)    0           batch_normalization_179[0][0]    
__________________________________________________________________________________________________
activation_182 (Activation)     (None, 4, 4, 256)    0           batch_normalization_182[0][0]    
__________________________________________________________________________________________________
block8_5_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_179[0][0]             
                                                                 activation_182[0][0]             
__________________________________________________________________________________________________
block8_5_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_5_mixed[0][0]             
__________________________________________________________________________________________________
block8_5 (Lambda)               (None, 4, 4, 2080)   0           block8_4_ac[0][0]                
                                                                 block8_5_conv[0][0]              
__________________________________________________________________________________________________
block8_5_ac (Activation)        (None, 4, 4, 2080)   0           block8_5[0][0]                   
__________________________________________________________________________________________________
conv2d_193 (Conv2D)             (None, 4, 4, 192)    399360      block8_5_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_184 (BatchN (None, 4, 4, 192)    576         conv2d_193[0][0]                 
__________________________________________________________________________________________________
activation_184 (Activation)     (None, 4, 4, 192)    0           batch_normalization_184[0][0]    
__________________________________________________________________________________________________
conv2d_194 (Conv2D)             (None, 4, 4, 224)    129024      activation_184[0][0]             
__________________________________________________________________________________________________
batch_normalization_185 (BatchN (None, 4, 4, 224)    672         conv2d_194[0][0]                 
__________________________________________________________________________________________________
activation_185 (Activation)     (None, 4, 4, 224)    0           batch_normalization_185[0][0]    
__________________________________________________________________________________________________
conv2d_192 (Conv2D)             (None, 4, 4, 192)    399360      block8_5_ac[0][0]                
__________________________________________________________________________________________________
conv2d_195 (Conv2D)             (None, 4, 4, 256)    172032      activation_185[0][0]             
__________________________________________________________________________________________________
batch_normalization_183 (BatchN (None, 4, 4, 192)    576         conv2d_192[0][0]                 
__________________________________________________________________________________________________
batch_normalization_186 (BatchN (None, 4, 4, 256)    768         conv2d_195[0][0]                 
__________________________________________________________________________________________________
activation_183 (Activation)     (None, 4, 4, 192)    0           batch_normalization_183[0][0]    
__________________________________________________________________________________________________
activation_186 (Activation)     (None, 4, 4, 256)    0           batch_normalization_186[0][0]    
__________________________________________________________________________________________________
block8_6_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_183[0][0]             
                                                                 activation_186[0][0]             
__________________________________________________________________________________________________
block8_6_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_6_mixed[0][0]             
__________________________________________________________________________________________________
block8_6 (Lambda)               (None, 4, 4, 2080)   0           block8_5_ac[0][0]                
                                                                 block8_6_conv[0][0]              
__________________________________________________________________________________________________
block8_6_ac (Activation)        (None, 4, 4, 2080)   0           block8_6[0][0]                   
__________________________________________________________________________________________________
conv2d_197 (Conv2D)             (None, 4, 4, 192)    399360      block8_6_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_188 (BatchN (None, 4, 4, 192)    576         conv2d_197[0][0]                 
__________________________________________________________________________________________________
activation_188 (Activation)     (None, 4, 4, 192)    0           batch_normalization_188[0][0]    
__________________________________________________________________________________________________
conv2d_198 (Conv2D)             (None, 4, 4, 224)    129024      activation_188[0][0]             
__________________________________________________________________________________________________
batch_normalization_189 (BatchN (None, 4, 4, 224)    672         conv2d_198[0][0]                 
__________________________________________________________________________________________________
activation_189 (Activation)     (None, 4, 4, 224)    0           batch_normalization_189[0][0]    
__________________________________________________________________________________________________
conv2d_196 (Conv2D)             (None, 4, 4, 192)    399360      block8_6_ac[0][0]                
__________________________________________________________________________________________________
conv2d_199 (Conv2D)             (None, 4, 4, 256)    172032      activation_189[0][0]             
__________________________________________________________________________________________________
batch_normalization_187 (BatchN (None, 4, 4, 192)    576         conv2d_196[0][0]                 
__________________________________________________________________________________________________
batch_normalization_190 (BatchN (None, 4, 4, 256)    768         conv2d_199[0][0]                 
__________________________________________________________________________________________________
activation_187 (Activation)     (None, 4, 4, 192)    0           batch_normalization_187[0][0]    
__________________________________________________________________________________________________
activation_190 (Activation)     (None, 4, 4, 256)    0           batch_normalization_190[0][0]    
__________________________________________________________________________________________________
block8_7_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_187[0][0]             
                                                                 activation_190[0][0]             
__________________________________________________________________________________________________
block8_7_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_7_mixed[0][0]             
__________________________________________________________________________________________________
block8_7 (Lambda)               (None, 4, 4, 2080)   0           block8_6_ac[0][0]                
                                                                 block8_7_conv[0][0]              
__________________________________________________________________________________________________
block8_7_ac (Activation)        (None, 4, 4, 2080)   0           block8_7[0][0]                   
__________________________________________________________________________________________________
conv2d_201 (Conv2D)             (None, 4, 4, 192)    399360      block8_7_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_192 (BatchN (None, 4, 4, 192)    576         conv2d_201[0][0]                 
__________________________________________________________________________________________________
activation_192 (Activation)     (None, 4, 4, 192)    0           batch_normalization_192[0][0]    
__________________________________________________________________________________________________
conv2d_202 (Conv2D)             (None, 4, 4, 224)    129024      activation_192[0][0]             
__________________________________________________________________________________________________
batch_normalization_193 (BatchN (None, 4, 4, 224)    672         conv2d_202[0][0]                 
__________________________________________________________________________________________________
activation_193 (Activation)     (None, 4, 4, 224)    0           batch_normalization_193[0][0]    
__________________________________________________________________________________________________
conv2d_200 (Conv2D)             (None, 4, 4, 192)    399360      block8_7_ac[0][0]                
__________________________________________________________________________________________________
conv2d_203 (Conv2D)             (None, 4, 4, 256)    172032      activation_193[0][0]             
__________________________________________________________________________________________________
batch_normalization_191 (BatchN (None, 4, 4, 192)    576         conv2d_200[0][0]                 
__________________________________________________________________________________________________
batch_normalization_194 (BatchN (None, 4, 4, 256)    768         conv2d_203[0][0]                 
__________________________________________________________________________________________________
activation_191 (Activation)     (None, 4, 4, 192)    0           batch_normalization_191[0][0]    
__________________________________________________________________________________________________
activation_194 (Activation)     (None, 4, 4, 256)    0           batch_normalization_194[0][0]    
__________________________________________________________________________________________________
block8_8_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_191[0][0]             
                                                                 activation_194[0][0]             
__________________________________________________________________________________________________
block8_8_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_8_mixed[0][0]             
__________________________________________________________________________________________________
block8_8 (Lambda)               (None, 4, 4, 2080)   0           block8_7_ac[0][0]                
                                                                 block8_8_conv[0][0]              
__________________________________________________________________________________________________
block8_8_ac (Activation)        (None, 4, 4, 2080)   0           block8_8[0][0]                   
__________________________________________________________________________________________________
conv2d_205 (Conv2D)             (None, 4, 4, 192)    399360      block8_8_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_196 (BatchN (None, 4, 4, 192)    576         conv2d_205[0][0]                 
__________________________________________________________________________________________________
activation_196 (Activation)     (None, 4, 4, 192)    0           batch_normalization_196[0][0]    
__________________________________________________________________________________________________
conv2d_206 (Conv2D)             (None, 4, 4, 224)    129024      activation_196[0][0]             
__________________________________________________________________________________________________
batch_normalization_197 (BatchN (None, 4, 4, 224)    672         conv2d_206[0][0]                 
__________________________________________________________________________________________________
activation_197 (Activation)     (None, 4, 4, 224)    0           batch_normalization_197[0][0]    
__________________________________________________________________________________________________
conv2d_204 (Conv2D)             (None, 4, 4, 192)    399360      block8_8_ac[0][0]                
__________________________________________________________________________________________________
conv2d_207 (Conv2D)             (None, 4, 4, 256)    172032      activation_197[0][0]             
__________________________________________________________________________________________________
batch_normalization_195 (BatchN (None, 4, 4, 192)    576         conv2d_204[0][0]                 
__________________________________________________________________________________________________
batch_normalization_198 (BatchN (None, 4, 4, 256)    768         conv2d_207[0][0]                 
__________________________________________________________________________________________________
activation_195 (Activation)     (None, 4, 4, 192)    0           batch_normalization_195[0][0]    
__________________________________________________________________________________________________
activation_198 (Activation)     (None, 4, 4, 256)    0           batch_normalization_198[0][0]    
__________________________________________________________________________________________________
block8_9_mixed (Concatenate)    (None, 4, 4, 448)    0           activation_195[0][0]             
                                                                 activation_198[0][0]             
__________________________________________________________________________________________________
block8_9_conv (Conv2D)          (None, 4, 4, 2080)   933920      block8_9_mixed[0][0]             
__________________________________________________________________________________________________
block8_9 (Lambda)               (None, 4, 4, 2080)   0           block8_8_ac[0][0]                
                                                                 block8_9_conv[0][0]              
__________________________________________________________________________________________________
block8_9_ac (Activation)        (None, 4, 4, 2080)   0           block8_9[0][0]                   
__________________________________________________________________________________________________
conv2d_209 (Conv2D)             (None, 4, 4, 192)    399360      block8_9_ac[0][0]                
__________________________________________________________________________________________________
batch_normalization_200 (BatchN (None, 4, 4, 192)    576         conv2d_209[0][0]                 
__________________________________________________________________________________________________
activation_200 (Activation)     (None, 4, 4, 192)    0           batch_normalization_200[0][0]    
__________________________________________________________________________________________________
conv2d_210 (Conv2D)             (None, 4, 4, 224)    129024      activation_200[0][0]             
__________________________________________________________________________________________________
batch_normalization_201 (BatchN (None, 4, 4, 224)    672         conv2d_210[0][0]                 
__________________________________________________________________________________________________
activation_201 (Activation)     (None, 4, 4, 224)    0           batch_normalization_201[0][0]    
__________________________________________________________________________________________________
conv2d_208 (Conv2D)             (None, 4, 4, 192)    399360      block8_9_ac[0][0]                
__________________________________________________________________________________________________
conv2d_211 (Conv2D)             (None, 4, 4, 256)    172032      activation_201[0][0]             
__________________________________________________________________________________________________
batch_normalization_199 (BatchN (None, 4, 4, 192)    576         conv2d_208[0][0]                 
__________________________________________________________________________________________________
batch_normalization_202 (BatchN (None, 4, 4, 256)    768         conv2d_211[0][0]                 
__________________________________________________________________________________________________
activation_199 (Activation)     (None, 4, 4, 192)    0           batch_normalization_199[0][0]    
__________________________________________________________________________________________________
activation_202 (Activation)     (None, 4, 4, 256)    0           batch_normalization_202[0][0]    
__________________________________________________________________________________________________
block8_10_mixed (Concatenate)   (None, 4, 4, 448)    0           activation_199[0][0]             
                                                                 activation_202[0][0]             
__________________________________________________________________________________________________
block8_10_conv (Conv2D)         (None, 4, 4, 2080)   933920      block8_10_mixed[0][0]            
__________________________________________________________________________________________________
block8_10 (Lambda)              (None, 4, 4, 2080)   0           block8_9_ac[0][0]                
                                                                 block8_10_conv[0][0]             
__________________________________________________________________________________________________
conv_7b (Conv2D)                (None, 4, 4, 1536)   3194880     block8_10[0][0]                  
__________________________________________________________________________________________________
conv_7b_bn (BatchNormalization) (None, 4, 4, 1536)   4608        conv_7b[0][0]                    
__________________________________________________________________________________________________
conv_7b_ac (Activation)         (None, 4, 4, 1536)   0           conv_7b_bn[0][0]                 
==================================================================================================
Total params: 54,336,736
Trainable params: 0
Non-trainable params: 54,336,736
__________________________________________________________________________________________________

Let's create a new model on top of the pretrained network.

In [ ]:

# Define the input shape
inputs = tf.keras.Input(shape=(180, 180, 3))

# stack the inputs to pretrained model and set training to false
x = pretrained_base_model(inputs, training=False)

# Add a pooling layer
x = tf.keras.layers.GlobalAveragePooling2D()(x)

# Add drop out layer

x = tf.keras.layers.Dropout(0.4)(x)

# Last output dense layer with 1 unit and sigmoid
output = tf.keras.layers.Dense(1, activation='sigmoid')(x)

# Build a model
model_3 = tf.keras.Model(inputs, output)

# Define the input shape inputs = tf.keras.Input(shape=(180, 180, 3)) # stack the inputs to pretrained model and set training to false x = pretrained_base_model(inputs, training=False) # Add a pooling layer x = tf.keras.layers.GlobalAveragePooling2D()(x) # Add drop out layer x = tf.keras.layers.Dropout(0.4)(x) # Last output dense layer with 1 unit and sigmoid output = tf.keras.layers.Dense(1, activation='sigmoid')(x) # Build a model model_3 = tf.keras.Model(inputs, output)

Now let's compile the model and train on our augmented data.

In [ ]:

model_3.compile(
    optimizer='rmsprop',
    loss='binary_crossentropy',
    metrics='accuracy'

)

model_3.compile( optimizer='rmsprop', loss='binary_crossentropy', metrics='accuracy' )

In [ ]:

history_3 = model_3.fit(
      train_generator,
      steps_per_epoch=train_steps,  
      epochs=25,
      validation_data=val_generator,
      validation_steps=val_steps)

history_3 = model_3.fit( train_generator, steps_per_epoch=train_steps, epochs=25, validation_data=val_generator, validation_steps=val_steps)

Epoch 1/25
100/100 [==============================] - 55s 412ms/step - loss: 0.2161 - accuracy: 0.9185 - val_loss: 0.0763 - val_accuracy: 0.9730
Epoch 2/25
100/100 [==============================] - 37s 372ms/step - loss: 0.1294 - accuracy: 0.9485 - val_loss: 0.0546 - val_accuracy: 0.9840
Epoch 3/25
100/100 [==============================] - 37s 369ms/step - loss: 0.1255 - accuracy: 0.9535 - val_loss: 0.0502 - val_accuracy: 0.9810
Epoch 4/25
100/100 [==============================] - 37s 372ms/step - loss: 0.1168 - accuracy: 0.9590 - val_loss: 0.0553 - val_accuracy: 0.9810
Epoch 5/25
100/100 [==============================] - 38s 376ms/step - loss: 0.1173 - accuracy: 0.9615 - val_loss: 0.0504 - val_accuracy: 0.9850
Epoch 6/25
100/100 [==============================] - 37s 373ms/step - loss: 0.1046 - accuracy: 0.9610 - val_loss: 0.0433 - val_accuracy: 0.9840
Epoch 7/25
100/100 [==============================] - 37s 374ms/step - loss: 0.1292 - accuracy: 0.9545 - val_loss: 0.0440 - val_accuracy: 0.9860
Epoch 8/25
100/100 [==============================] - 37s 372ms/step - loss: 0.1085 - accuracy: 0.9605 - val_loss: 0.0384 - val_accuracy: 0.9850
Epoch 9/25
100/100 [==============================] - 37s 371ms/step - loss: 0.1075 - accuracy: 0.9650 - val_loss: 0.0364 - val_accuracy: 0.9890
Epoch 10/25
100/100 [==============================] - 37s 371ms/step - loss: 0.0975 - accuracy: 0.9625 - val_loss: 0.0386 - val_accuracy: 0.9850
Epoch 11/25
100/100 [==============================] - 37s 369ms/step - loss: 0.1035 - accuracy: 0.9655 - val_loss: 0.0433 - val_accuracy: 0.9800
Epoch 12/25
100/100 [==============================] - 37s 374ms/step - loss: 0.0999 - accuracy: 0.9695 - val_loss: 0.0503 - val_accuracy: 0.9810
Epoch 13/25
100/100 [==============================] - 37s 373ms/step - loss: 0.0930 - accuracy: 0.9650 - val_loss: 0.0480 - val_accuracy: 0.9860
Epoch 14/25
100/100 [==============================] - 38s 375ms/step - loss: 0.0849 - accuracy: 0.9710 - val_loss: 0.0417 - val_accuracy: 0.9870
Epoch 15/25
100/100 [==============================] - 38s 376ms/step - loss: 0.0993 - accuracy: 0.9645 - val_loss: 0.0439 - val_accuracy: 0.9840
Epoch 16/25
100/100 [==============================] - 37s 374ms/step - loss: 0.0887 - accuracy: 0.9715 - val_loss: 0.0418 - val_accuracy: 0.9890
Epoch 17/25
100/100 [==============================] - 37s 372ms/step - loss: 0.1006 - accuracy: 0.9700 - val_loss: 0.0405 - val_accuracy: 0.9900
Epoch 18/25
100/100 [==============================] - 37s 372ms/step - loss: 0.0873 - accuracy: 0.9690 - val_loss: 0.0406 - val_accuracy: 0.9860
Epoch 19/25
100/100 [==============================] - 37s 371ms/step - loss: 0.1124 - accuracy: 0.9615 - val_loss: 0.0408 - val_accuracy: 0.9850
Epoch 20/25
100/100 [==============================] - 37s 371ms/step - loss: 0.0999 - accuracy: 0.9645 - val_loss: 0.0448 - val_accuracy: 0.9870
Epoch 21/25
100/100 [==============================] - 38s 376ms/step - loss: 0.0922 - accuracy: 0.9695 - val_loss: 0.0466 - val_accuracy: 0.9860
Epoch 22/25
100/100 [==============================] - 37s 371ms/step - loss: 0.0859 - accuracy: 0.9685 - val_loss: 0.0393 - val_accuracy: 0.9870
Epoch 23/25
100/100 [==============================] - 37s 372ms/step - loss: 0.1002 - accuracy: 0.9650 - val_loss: 0.0628 - val_accuracy: 0.9820
Epoch 24/25
100/100 [==============================] - 37s 370ms/step - loss: 0.1020 - accuracy: 0.9635 - val_loss: 0.0428 - val_accuracy: 0.9850
Epoch 25/25
100/100 [==============================] - 37s 369ms/step - loss: 0.0784 - accuracy: 0.9705 - val_loss: 0.0408 - val_accuracy: 0.9850

This is soo impressive, over 97% accuracy on training. You can see we didn't have to even train for 25 epochs, 10 would give great results. Let's plot the results and also later change the epochs to 10.

In [ ]:

model_history_3 = history_3.history

acc = model_history_3['accuracy']
val_acc = model_history_3['val_accuracy']
loss = model_history_3['loss']
val_loss = model_history_3['val_loss']

epochs = history_3.epoch

plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

model_history_3 = history_3.history acc = model_history_3['accuracy'] val_acc = model_history_3['val_accuracy'] loss = model_history_3['loss'] val_loss = model_history_3['val_loss'] epochs = history_3.epoch plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

<Figure size 432x288 with 0 Axes>

We have said that using a pretrained model is like standing on the shoulder of the giant. And that's pretty clear on the above graph, the accuracy started at ~90%.

N.B. Calling model.fit(...) continue where the training left. To avoid adding up on previous training, you have to redefine & compile the model.

In [ ]:

# Define the input shape
inputs = tf.keras.Input(shape=(180, 180, 3))

# stack the inputs to pretrained model and set training to false
x = pretrained_base_model(inputs, training=False)

# Add a pooling layer
x = tf.keras.layers.GlobalAveragePooling2D()(x)

# Add drop out layer

x = tf.keras.layers.Dropout(0.4)(x)

# Last output dense layer with 1 unit and sigmoid
output = tf.keras.layers.Dense(1, activation='sigmoid')(x)

# Build a model
model_4 = tf.keras.Model(inputs, output)

# Define the input shape inputs = tf.keras.Input(shape=(180, 180, 3)) # stack the inputs to pretrained model and set training to false x = pretrained_base_model(inputs, training=False) # Add a pooling layer x = tf.keras.layers.GlobalAveragePooling2D()(x) # Add drop out layer x = tf.keras.layers.Dropout(0.4)(x) # Last output dense layer with 1 unit and sigmoid output = tf.keras.layers.Dense(1, activation='sigmoid')(x) # Build a model model_4 = tf.keras.Model(inputs, output)

In [ ]:

model_4.compile(
    optimizer='rmsprop',
    loss='binary_crossentropy',
    metrics='accuracy'

)

model_4.compile( optimizer='rmsprop', loss='binary_crossentropy', metrics='accuracy' )

In [ ]:

# Train for 10 epochs

history_4 = model_4.fit(
      train_generator,
      steps_per_epoch=train_steps,  
      epochs=10,
      validation_data=val_generator,
      validation_steps=val_steps)

# Train for 10 epochs history_4 = model_4.fit( train_generator, steps_per_epoch=train_steps, epochs=10, validation_data=val_generator, validation_steps=val_steps)

Epoch 1/10
100/100 [==============================] - 53s 409ms/step - loss: 0.2150 - accuracy: 0.9175 - val_loss: 0.0557 - val_accuracy: 0.9830
Epoch 2/10
100/100 [==============================] - 37s 369ms/step - loss: 0.1386 - accuracy: 0.9460 - val_loss: 0.0491 - val_accuracy: 0.9830
Epoch 3/10
100/100 [==============================] - 37s 369ms/step - loss: 0.1041 - accuracy: 0.9595 - val_loss: 0.0511 - val_accuracy: 0.9830
Epoch 4/10
100/100 [==============================] - 37s 368ms/step - loss: 0.1119 - accuracy: 0.9610 - val_loss: 0.0522 - val_accuracy: 0.9820
Epoch 5/10
100/100 [==============================] - 37s 368ms/step - loss: 0.1112 - accuracy: 0.9615 - val_loss: 0.0539 - val_accuracy: 0.9810
Epoch 6/10
100/100 [==============================] - 37s 371ms/step - loss: 0.1100 - accuracy: 0.9645 - val_loss: 0.0483 - val_accuracy: 0.9820
Epoch 7/10
100/100 [==============================] - 37s 372ms/step - loss: 0.0999 - accuracy: 0.9615 - val_loss: 0.0463 - val_accuracy: 0.9840
Epoch 8/10
100/100 [==============================] - 37s 372ms/step - loss: 0.0949 - accuracy: 0.9610 - val_loss: 0.0455 - val_accuracy: 0.9860
Epoch 9/10
100/100 [==============================] - 37s 370ms/step - loss: 0.1097 - accuracy: 0.9620 - val_loss: 0.0512 - val_accuracy: 0.9840
Epoch 10/10
100/100 [==============================] - 37s 371ms/step - loss: 0.0996 - accuracy: 0.9600 - val_loss: 0.0425 - val_accuracy: 0.9870

In [ ]:

model_history_4 = history_4.history

acc = model_history_4['accuracy']
val_acc = model_history_4['val_accuracy']
loss = model_history_4['loss']
val_loss = model_history_4['val_loss']

epochs = history_4.epoch

plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

model_history_4 = history_4.history acc = model_history_4['accuracy'] val_acc = model_history_4['val_accuracy'] loss = model_history_4['loss'] val_loss = model_history_4['val_loss'] epochs = history_4.epoch plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

<Figure size 432x288 with 0 Axes>

The reason why validation accuracy is better than training accuracy, or why validation loss is less than training loss, it's because the pretrained model has Batch Normalization and dropout layers. These regularization layers affects the accuracy during training, but they are turned off during validation.

Also, during the training, accuracy and loss are averaged per epoch, while during validation phase, accuracy and loss are computed on a model that has already trained longer.

You can notice this if you keep the eye on progress bar during the training. The training metrics change per step/epoch, and at the end of the epoch, the average loss/accuracy are reported. On the other hand, the validation metrics are computed after training on each epoch.

3.7 Saving, Loading, and Testing a Model on New Images¶

In [ ]:

# Saving a model

model_4.save('final_model.h5')

# or 

#tf.keras.models.save_model(model_4, 'final_model.h5')

# Saving a model model_4.save('final_model.h5') # or #tf.keras.models.save_model(model_4, 'final_model.h5')

In [ ]:

# load a saved model

loaded_model = tf.keras.models.load_model('/content/final_model.h5')

# load a saved model loaded_model = tf.keras.models.load_model('/content/final_model.h5')

In [ ]:

# function to download some images from the internet

def get_image(image_name,url):

  image = tf.keras.utils.get_file(image_name,url)

  return image

# function to download some images from the internet def get_image(image_name,url): image = tf.keras.utils.get_file(image_name,url) return image

In [ ]:

# Urls of cats and dogs images

cat_1 = 'https://upload.wikimedia.org/wikipedia/commons/b/b6/Felis_catus-cat_on_snow.jpg'
cat_2 = 'https://upload.wikimedia.org/wikipedia/commons/6/69/June_odd-eyed-cat_cropped.jpg'

dog_1 = 'https://upload.wikimedia.org/wikipedia/commons/b/b5/2008-08-28_White_German_Shepherd_ready.jpg'
dog_2 = 'https://upload.wikimedia.org/wikipedia/commons/d/df/Dogs_were_not_around_in_the_early_Cenozoic.png'

# Urls of cats and dogs images cat_1 = 'https://upload.wikimedia.org/wikipedia/commons/b/b6/Felis_catus-cat_on_snow.jpg' cat_2 = 'https://upload.wikimedia.org/wikipedia/commons/6/69/June_odd-eyed-cat_cropped.jpg' dog_1 = 'https://upload.wikimedia.org/wikipedia/commons/b/b5/2008-08-28_White_German_Shepherd_ready.jpg' dog_2 = 'https://upload.wikimedia.org/wikipedia/commons/d/df/Dogs_were_not_around_in_the_early_Cenozoic.png'

In [ ]:

cat_im1 = get_image('cat_1',cat_1)
cat_im2 = get_image('cat_2',cat_2)

dog_im1 = get_image('dog_1',dog_1)
dog_im2 = get_image('dog_2',dog_2)

cat_im1 = get_image('cat_1',cat_1) cat_im2 = get_image('cat_2',cat_2) dog_im1 = get_image('dog_1',dog_1) dog_im2 = get_image('dog_2',dog_2)

Downloading data from https://upload.wikimedia.org/wikipedia/commons/b/b6/Felis_catus-cat_on_snow.jpg
2129920/2125399 [==============================] - 0s 0us/step
2138112/2125399 [==============================] - 0s 0us/step
Downloading data from https://upload.wikimedia.org/wikipedia/commons/6/69/June_odd-eyed-cat_cropped.jpg
581632/580625 [==============================] - 0s 0us/step
589824/580625 [==============================] - 0s 0us/step
Downloading data from https://upload.wikimedia.org/wikipedia/commons/b/b5/2008-08-28_White_German_Shepherd_ready.jpg
4276224/4271571 [==============================] - 0s 0us/step
4284416/4271571 [==============================] - 0s 0us/step
Downloading data from https://upload.wikimedia.org/wikipedia/commons/d/df/Dogs_were_not_around_in_the_early_Cenozoic.png
1966080/1959233 [==============================] - 0s 0us/step
1974272/1959233 [==============================] - 0s 0us/step

In [ ]:

def predict(model, image):
  ''' Take model & image, preprocess image, make predictions, and return results'''

  image = tf.keras.preprocessing.image.load_img(image, target_size=(180,180))
  image = tf.keras.preprocessing.image.img_to_array(image)
  image = image/255.0
  image = tf.expand_dims(image, 0)

  predicted_class = model.predict(image)[0]
  rounded_predicted_class = tf.round(predicted_class)

  print(f"DISPLAYING THE PREDICTION RESULTS:\n \
    ---------- \n \
    Predicted class probability: {predicted_class}\n \
    Predicted class: {rounded_predicted_class}\n \
    Predicted class name: {'Dog' if rounded_predicted_class == 1 else 'Cat'}")

def predict(model, image): ''' Take model & image, preprocess image, make predictions, and return results''' image = tf.keras.preprocessing.image.load_img(image, target_size=(180,180)) image = tf.keras.preprocessing.image.img_to_array(image) image = image/255.0 image = tf.expand_dims(image, 0) predicted_class = model.predict(image)[0] rounded_predicted_class = tf.round(predicted_class) print(f"DISPLAYING THE PREDICTION RESULTS:\n \ ---------- \n \ Predicted class probability: {predicted_class}\n \ Predicted class: {rounded_predicted_class}\n \ Predicted class name: {'Dog' if rounded_predicted_class == 1 else 'Cat'}")

In [ ]:

predict(loaded_model, dog_im1)
plt.imshow(tf.keras.preprocessing.image.load_img(dog_im1, target_size=(180,180)));

predict(loaded_model, dog_im1) plt.imshow(tf.keras.preprocessing.image.load_img(dog_im1, target_size=(180,180)));

DISPLAYING THE PREDICTION RESULTS:
     ---------- 
     Predicted class probability: [0.99823844]
     Predicted class: [1.]
     Predicted class name: Dog

In [ ]:

predict(loaded_model, cat_im1)
plt.imshow(tf.keras.preprocessing.image.load_img(cat_im1, target_size=(180,180)));

predict(loaded_model, cat_im1) plt.imshow(tf.keras.preprocessing.image.load_img(cat_im1, target_size=(180,180)));

DISPLAYING THE PREDICTION RESULTS:
     ---------- 
     Predicted class probability: [0.00061188]
     Predicted class: [0.]
     Predicted class name: Cat

In [ ]:

predict(loaded_model, dog_im2)
plt.imshow(tf.keras.preprocessing.image.load_img(dog_im2, target_size=(180,180)));

predict(loaded_model, dog_im2) plt.imshow(tf.keras.preprocessing.image.load_img(dog_im2, target_size=(180,180)));

DISPLAYING THE PREDICTION RESULTS:
     ---------- 
     Predicted class probability: [0.9991021]
     Predicted class: [1.]
     Predicted class name: Dog

In [ ]:

predict(loaded_model, cat_im2)
plt.imshow(tf.keras.preprocessing.image.load_img(cat_im2, target_size=(180,180)));

predict(loaded_model, cat_im2) plt.imshow(tf.keras.preprocessing.image.load_img(cat_im2, target_size=(180,180)));

DISPLAYING THE PREDICTION RESULTS:
     ---------- 
     Predicted class probability: [2.567915e-06]
     Predicted class: [0.]
     Predicted class name: Cat

4. Image Augmentation with Keras Image Augmentation Layers¶

ImageDataGenerator makes it easy to augment images while while loading them from the directory at the same time.

But there is a time you may want the augmentation to takes place inside the model. In that case, you can use Keras Image Augmentation layers. These layers can be part of the model, and what that means is that the input image to a model will be augmented first before undergoing further preprocessing.

But why would you want the images to be augmented inside the model? The single most advantage of it is that when you save a model, these layers are saved too. If you deploy a model to the mobile device for example, the images will be preprocessed automatically, and you won't have to write custom preprocessing codes. Also, as they are together with other model layers, you will benefit from GPU acceleration.

These types of layers can also be used to preprocess and augment images directly, or outside the model.

A something worth noting is that just like dropout and batch normalization layers, data augmentation layers are only active during model training(model.fit()), and inactive during model evaluation(model.evaluate()) and predictions (model.predict())

Before TensorFlow 2.6, these layers were in experimentation. But in version 2.6, they can be used as other layers and there is no worries that they will go anywhere.

Here is a list of available image augmentation layers:

• RandomCrop layer
• RandomFlip layer
• RandomTranslation layer
• RandomRotation layer
• RandomZoom layer
• RandomHeight layer
• RandomWidth layer
• RandomContrast layer

We are going to practice how to use these layers inside and outside the model. We are also going to do something different: We will use TensorFlow Datasets.

4.1 Getting the Data Again¶

Let's use the same dataset (cats_vs_dogs), but we will get it from TensorFlow datasets to also learn how to work with TF datasets.

The cats_vs_dogs is available in TensorFlow datasets and it contains 23,262 images.

In [1]:

import tensorflow as tf
from tensorflow import keras
import tensorflow_datasets as tfds
import matplotlib.pyplot as plt
import numpy as np

import tensorflow as tf from tensorflow import keras import tensorflow_datasets as tfds import matplotlib.pyplot as plt import numpy as np

Here are few notes about loading data with tfds.load(...)

• The first thing to be provided while loading the dataset, it is the name of such dataset. The names are pretty clear in TFDS catalog.
• We can directly split the data into training and test with split argument. With split=('train[:80%]','train[80%:]'), 80% of train is assigned to training data, and 20% of it is allocated to validation data.

A quick note about these 3 popular sets: training set is used for training the model, validation set is for evaluating the model during training, and test set is a new data that we use to test the model. In the real world projects, validation set should have the same distribution as test set.

• Setting as_supervised to true will return a tuple of image and label(image, label)instead of a dictionary({'image': img, 'label': label}).

• And with_info=True will return the information of the concerned dataset.

In [2]:

(train_data, val_data), info = tfds.load('cats_vs_dogs:4.0.0',
                                         split=('train[:80%]','train[80%:]'),
                                         as_supervised=True,
                                         with_info = True
)

(train_data, val_data), info = tfds.load('cats_vs_dogs:4.0.0', split=('train[:80%]','train[80%:]'), as_supervised=True, with_info = True )

Downloading and preparing dataset cats_vs_dogs/4.0.0 (download: 786.68 MiB, generated: Unknown size, total: 786.68 MiB) to /root/tensorflow_datasets/cats_vs_dogs/4.0.0...

WARNING:absl:1738 images were corrupted and were skipped

Shuffling and writing examples to /root/tensorflow_datasets/cats_vs_dogs/4.0.0.incompleteRB4W93/cats_vs_dogs-train.tfrecord

Dataset cats_vs_dogs downloaded and prepared to /root/tensorflow_datasets/cats_vs_dogs/4.0.0. Subsequent calls will reuse this data.

We can see the dataset size by...

In [3]:

print("The number images in training set: {}".format(len(train_data)))
print("The number images in validation set: {}".format(len(val_data)))

print("The number images in training set: {}".format(len(train_data))) print("The number images in validation set: {}".format(len(val_data)))

The number images in training set: 18610
The number images in validation set: 4652

Or use info to get number of examples.

In [4]:

# Getting the number of examples 

info.splits['train'].num_examples

# Getting the number of examples info.splits['train'].num_examples

Out[4]:

23262

Let's also get class names from info.features

In [5]:

# getting class names
# Only display the first 10 classes..There are 120 classes

class_names = info.features['label'].names
class_names

# getting class names # Only display the first 10 classes..There are 120 classes class_names = info.features['label'].names class_names

Out[5]:

['cat', 'dog']

In [6]:

# Number of classes

info.features['label'].num_classes

# Number of classes info.features['label'].num_classes

Out[6]:

2

Now that the data is loaded, let's visualize some images. It's always a cool thing to do.

4.2 Looking in Images¶

As always, it's good to peep in images and see if there are images that are labelled incorrectly, or have an extenstion of .pdf when in fact all images should be .png or jpg.

As you might guess, TensorFlow datasets are remarkably prepared, but in real life, incorrect labelling and formats can exist.

Let's use tfds.visualization(..) to visualize some images.

In [7]:

fig = tfds.show_examples(train_data, info)

fig = tfds.show_examples(train_data, info)

4.3 Preparing Data: Building a Training Pipeline¶

Now, let's prepare the data, specifically, building an input pipeline.

First, the tensorflow datasets are returned as uint8, we need to resize and normalize the images while also converting it to float32.

In [8]:

def preprocess(image, label):

  """ 
  Take image and label, 
  Resize images to 180, 180
  convert images to float32, and return converted image &label
  """

  resized_image = tf.image.resize(image, [180, 180])
  normalized_img = tf.cast(resized_image, tf.float32)/255.0

  return normalized_img, label

def preprocess(image, label): """ Take image and label, Resize images to 180, 180 convert images to float32, and return converted image &label """ resized_image = tf.image.resize(image, [180, 180]) normalized_img = tf.cast(resized_image, tf.float32)/255.0 return normalized_img, label

Now, we can apply the above function to the dataset with map. Along the way, we will shuffle it, and batch the images.

We do not shuffle the test set.

In [9]:

def train_data_prep(data, shuffle_size, batch_size):

  data = data.map(preprocess)
  data = data.cache()
  data = data.shuffle(shuffle_size)
  data = data.batch(batch_size)
  data = data.prefetch(1)

  return data

def val_data_prep(data, batch_size):

  data = data.map(preprocess)
  data = data.batch(batch_size)
  data = data.cache()
  data = data.prefetch(1)

  return data

def train_data_prep(data, shuffle_size, batch_size): data = data.map(preprocess) data = data.cache() data = data.shuffle(shuffle_size) data = data.batch(batch_size) data = data.prefetch(1) return data def val_data_prep(data, batch_size): data = data.map(preprocess) data = data.batch(batch_size) data = data.cache() data = data.prefetch(1) return data

In [11]:

train_data_prepared = train_data_prep(train_data, 1000, 32)
val_data_prepared = val_data_prep(val_data, 32)

train_data_prepared = train_data_prep(train_data, 1000, 32) val_data_prepared = val_data_prep(val_data, 32)

4.4 Augmenting Images with Augmentation Layers¶

Learn more about Image Augmentation layers at Keras doc.

In [12]:

augmentation_layers = tf.keras.Sequential([
     tf.keras.layers.RandomFlip('horizontal_and_vertical'),
     tf.keras.layers.RandomRotation(0.2),
     tf.keras.layers.RandomContrast(0.4)
])

augmentation_layers = tf.keras.Sequential([ tf.keras.layers.RandomFlip('horizontal_and_vertical'), tf.keras.layers.RandomRotation(0.2), tf.keras.layers.RandomContrast(0.4) ])

To illustrate augmentation, let's apply the above layers to the images. Because the images are already converted into batch, running the cell below will plot a given image in a batch (running it again displays a different image). And it will later be augmented.

In [ ]:

image = train_data_prepared.take(1)
image, label = next(iter(image))

plt.imshow(image[0])
plt.title(class_names[label[0]]);

image = train_data_prepared.take(1) image, label = next(iter(image)) plt.imshow(image[0]) plt.title(class_names[label[0]]);

In [ ]:

plt.figure(figsize=(12,12))

for i in range(16):
  augmented_image = augmentation_layers(image)
  ax = plt.subplot(4,4,i+1);
  plt.imshow(augmented_image[0]);
  plt.axis('off');

plt.figure(figsize=(12,12)) for i in range(16): augmented_image = augmentation_layers(image) ax = plt.subplot(4,4,i+1); plt.imshow(augmented_image[0]); plt.axis('off');

WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

As we can see from the above, it is pretty clear that the input image were transformed into many ways, from flipping, contrast change to being cropped. It's very amazing to just stack layers and use them to augment images on the fly.

4.5 First Option: Using the Augmentation layers Outside the Model¶

Like we have seen, the augmentation layers can be used inside the model or outside the model. To use them outide the model is pretty straight. Just use a lambda function to get the image and label and apply the augmentation layers to every single image with a map function and return augmented image and label.

Note that we do not augment the validation and test data.

In [ ]:

train_data_augmented = train_data_prepared.map(lambda image, label: (augmentation_layers(image, training=True), label))

train_data_augmented = train_data_prepared.map(lambda image, label: (augmentation_layers(image, training=True), label))

Now we can build the Convnet model.

In [ ]:

model = tf.keras.Sequential([
                             
  # First Conv and maxpooling layer                         
  tf.keras.layers.Conv2D(32, 3, padding='same', activation='relu'),
  tf.keras.layers.MaxPooling2D(2,2),

  # Second Conv and maxpooling layer 
  tf.keras.layers.Conv2D(64, 3, padding='same', activation='relu'),
  tf.keras.layers.MaxPooling2D(2,2),

  # Third Conv and maxpooling layer 
  tf.keras.layers.Conv2D(128, 3, padding='same', activation='relu'),
  tf.keras.layers.MaxPooling2D(2,2),
  
  # Fourth Conv and maxpooling layer 
  tf.keras.layers.Conv2D(256, 3, padding='same', activation='relu'),
  tf.keras.layers.MaxPooling2D(2,2),

  # Flattening the feature maps into a 1D vector 
  tf.keras.layers.Flatten(),

  # Adding fully connected layers
  tf.keras.layers.Dense(256, activation='relu'),
  
  # Applying dropout to avoid overfitting
  tf.keras.layers.Dropout(0.3),

  tf.keras.layers.Dense(1, activation='sigmoid')
])

model = tf.keras.Sequential([ # First Conv and maxpooling layer tf.keras.layers.Conv2D(32, 3, padding='same', activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Second Conv and maxpooling layer tf.keras.layers.Conv2D(64, 3, padding='same', activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Third Conv and maxpooling layer tf.keras.layers.Conv2D(128, 3, padding='same', activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Fourth Conv and maxpooling layer tf.keras.layers.Conv2D(256, 3, padding='same', activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Flattening the feature maps into a 1D vector tf.keras.layers.Flatten(), # Adding fully connected layers tf.keras.layers.Dense(256, activation='relu'), # Applying dropout to avoid overfitting tf.keras.layers.Dropout(0.3), tf.keras.layers.Dense(1, activation='sigmoid') ])

In [ ]:

model.compile(optimizer='adam',
              loss='binary_crossentropy',
              metrics=['accuracy'])

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

In [ ]:

history = model.fit(
  train_data_augmented,
  validation_data=val_data_prepared,
  epochs=10
)

history = model.fit( train_data_augmented, validation_data=val_data_prepared, epochs=10 )

Epoch 1/10
582/582 [==============================] - 1670s 3s/step - loss: 0.6910 - accuracy: 0.5494 - val_loss: 0.6827 - val_accuracy: 0.5363
Epoch 2/10
582/582 [==============================] - 1658s 3s/step - loss: 0.6480 - accuracy: 0.6238 - val_loss: 0.6178 - val_accuracy: 0.6492
Epoch 3/10
582/582 [==============================] - 1657s 3s/step - loss: 0.6048 - accuracy: 0.6724 - val_loss: 0.5553 - val_accuracy: 0.7104
Epoch 4/10
582/582 [==============================] - 1666s 3s/step - loss: 0.5841 - accuracy: 0.6908 - val_loss: 0.5539 - val_accuracy: 0.7279
Epoch 5/10
582/582 [==============================] - 1659s 3s/step - loss: 0.5639 - accuracy: 0.7081 - val_loss: 0.5472 - val_accuracy: 0.7317
Epoch 6/10
582/582 [==============================] - 1659s 3s/step - loss: 0.5424 - accuracy: 0.7246 - val_loss: 0.5120 - val_accuracy: 0.7569
Epoch 7/10
582/582 [==============================] - 1662s 3s/step - loss: 0.5366 - accuracy: 0.7311 - val_loss: 0.4940 - val_accuracy: 0.7614
Epoch 8/10
582/582 [==============================] - 1658s 3s/step - loss: 0.5133 - accuracy: 0.7475 - val_loss: 0.4917 - val_accuracy: 0.7607
Epoch 9/10
582/582 [==============================] - 1652s 3s/step - loss: 0.4938 - accuracy: 0.7596 - val_loss: 0.4888 - val_accuracy: 0.7661
Epoch 10/10
582/582 [==============================] - 1652s 3s/step - loss: 0.4806 - accuracy: 0.7707 - val_loss: 0.4297 - val_accuracy: 0.8048

In [ ]:

model_history_5 = history.history

acc = model_history_5['accuracy']
val_acc = model_history_5['val_accuracy']
loss = model_history_5['loss']
val_loss = model_history_5['val_loss']

epochs = history.epoch

plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

model_history_5 = history.history acc = model_history_5['accuracy'] val_acc = model_history_5['val_accuracy'] loss = model_history_5['loss'] val_loss = model_history_5['val_loss'] epochs = history.epoch plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

<Figure size 432x288 with 0 Axes>

That was about training the model with the augmented images. Like we said, we can also have the the augmentation layers inside the model and take the advantage of GPU accelerations.

4.6 Second Option: Using the Augmentation layers Inside the Model¶

The augmentation layers that we created early will be the first layer in the model. At every training epoch, a given image is augmented(as described in the layers we defined above) before undergoing further transformations.

In [13]:

model = tf.keras.Sequential([
                             
  # Augmentation layers

  augmentation_layers,
              
  # First Conv and maxpooling layer                         
  tf.keras.layers.Conv2D(32, 3, padding='same', activation='relu'),
  tf.keras.layers.MaxPooling2D(2,2),

  # Second Conv and maxpooling layer 
  tf.keras.layers.Conv2D(64, 3, padding='same', activation='relu'),
  tf.keras.layers.MaxPooling2D(2,2),

  # Third Conv and maxpooling layer 
  tf.keras.layers.Conv2D(128, 3, padding='same', activation='relu'),
  tf.keras.layers.MaxPooling2D(2,2),
  
  # Fourth Conv and maxpooling layer 
  tf.keras.layers.Conv2D(256, 3, padding='same', activation='relu'),
  tf.keras.layers.MaxPooling2D(2,2),

  # Flattening the feature maps into a 1D vector 
  tf.keras.layers.Flatten(),

  # Adding fully connected layers
  tf.keras.layers.Dense(256, activation='relu'),
  
  # Applying dropout to avoid overfitting
  tf.keras.layers.Dropout(0.3),

  tf.keras.layers.Dense(1, activation='sigmoid')
])

model = tf.keras.Sequential([ # Augmentation layers augmentation_layers, # First Conv and maxpooling layer tf.keras.layers.Conv2D(32, 3, padding='same', activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Second Conv and maxpooling layer tf.keras.layers.Conv2D(64, 3, padding='same', activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Third Conv and maxpooling layer tf.keras.layers.Conv2D(128, 3, padding='same', activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Fourth Conv and maxpooling layer tf.keras.layers.Conv2D(256, 3, padding='same', activation='relu'), tf.keras.layers.MaxPooling2D(2,2), # Flattening the feature maps into a 1D vector tf.keras.layers.Flatten(), # Adding fully connected layers tf.keras.layers.Dense(256, activation='relu'), # Applying dropout to avoid overfitting tf.keras.layers.Dropout(0.3), tf.keras.layers.Dense(1, activation='sigmoid') ])

In [14]:

model.compile(optimizer='adam',
              loss='binary_crossentropy',
              metrics=['accuracy'])

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

In [15]:

history = model.fit(
  train_data_prepared,
  validation_data=val_data_prepared,
  epochs=10
)

history = model.fit( train_data_prepared, validation_data=val_data_prepared, epochs=10 )

Epoch 1/10
582/582 [==============================] - 1711s 3s/step - loss: 0.6802 - accuracy: 0.5593 - val_loss: 0.6397 - val_accuracy: 0.6182
Epoch 2/10
582/582 [==============================] - 1662s 3s/step - loss: 0.6338 - accuracy: 0.6377 - val_loss: 0.6296 - val_accuracy: 0.6505
Epoch 3/10
582/582 [==============================] - 1656s 3s/step - loss: 0.5885 - accuracy: 0.6877 - val_loss: 0.6099 - val_accuracy: 0.6808
Epoch 4/10
582/582 [==============================] - 1657s 3s/step - loss: 0.5749 - accuracy: 0.7015 - val_loss: 0.5250 - val_accuracy: 0.7367
Epoch 5/10
582/582 [==============================] - 1660s 3s/step - loss: 0.5593 - accuracy: 0.7127 - val_loss: 0.5245 - val_accuracy: 0.7416
Epoch 6/10
582/582 [==============================] - 1655s 3s/step - loss: 0.5433 - accuracy: 0.7236 - val_loss: 0.5365 - val_accuracy: 0.7281
Epoch 7/10
582/582 [==============================] - 1652s 3s/step - loss: 0.5353 - accuracy: 0.7303 - val_loss: 0.4975 - val_accuracy: 0.7620
Epoch 8/10
582/582 [==============================] - 1658s 3s/step - loss: 0.5183 - accuracy: 0.7440 - val_loss: 0.4943 - val_accuracy: 0.7610
Epoch 9/10
582/582 [==============================] - 1659s 3s/step - loss: 0.5099 - accuracy: 0.7507 - val_loss: 0.4948 - val_accuracy: 0.7584
Epoch 10/10
582/582 [==============================] - 1645s 3s/step - loss: 0.4928 - accuracy: 0.7609 - val_loss: 0.4532 - val_accuracy: 0.7840

In [18]:

model_history_6 = history.history

acc = model_history_6['accuracy']
val_acc = model_history_6['val_accuracy']
loss = model_history_6['loss']
val_loss = model_history_6['val_loss']

epochs = history.epoch

plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

model_history_6 = history.history acc = model_history_6['accuracy'] val_acc = model_history_6['val_accuracy'] loss = model_history_6['loss'] val_loss = model_history_6['val_loss'] epochs = history.epoch plot_acc_loss(acc, val_acc, loss, val_loss, epochs)

<Figure size 432x288 with 0 Axes>

That's it for using the Keras image augmentation layers. As it looks, the model with the augmentation layers inside seems bstter than the one with pre augmented images. And they were all trained with similar hyperparameters.

5. Further Learning¶

The following are the most recommended courses to learn more about machine learning basics and computer vision

• Google Machine Learning Crash Course for foundations of Machine Learning

• Intro to Deep Learning MIT (Lecture 1 and 3) for quick foundations of Deep Learning and Deep Computer Vision

• Deep Learning Specialization, Andrew Ng. This is a great course to give you indepth foundations of Deep Learning. Free on YouTube.

• Fast.AI Practical Deep Learning for Coders. This is the best in class course (very practical and has a high community ratings).

BACK TO TOP¶

In [ ]: