NMZivkovic

history = googlenet.fit(data_loader.train_batches,
                    epochs=10,
                    validation_data=data_loader.validation_batches)

NMZivkovic

history = vgg16.fit(data_loader.train_batches,
                    epochs=10,
                    validation_data=data_loader.validation_batches)

NMZivkovic

steps_per_epoch = round(data_loader.num_train_examples)//BATCH_SIZE
validation_steps = 20

loss1, accuracy1 = vgg16.evaluate(data_loader.validation_batches, steps = 20)
loss2, accuracy2 = googlenet.evaluate(data_loader.validation_batches, steps = 20)
loss3, accuracy3 = resnet.evaluate(data_loader.validation_batches, steps = 20)

print("--------VGG16---------")
print("Initial loss: {:.2f}".format(loss1))
print("Initial accuracy: {:.2f}".format(accuracy1))

NMZivkovic

base_learning_rate = 0.0001

vgg16_base.trainable = False
vgg16 = Wrapper(vgg16_base)
vgg16.compile(optimizer=tf.keras.optimizers.RMSprop(lr=base_learning_rate),
              loss='binary_crossentropy',
              metrics=['accuracy'])

googlenet_base.trainable = False
googlenet = Wrapper(googlenet_base)

NMZivkovic

vgg16_base = tf.keras.applications.VGG16(input_shape=IMG_SHAPE, include_top=False, weights='imagenet')
googlenet_base = tf.keras.applications.InceptionV3(input_shape=IMG_SHAPE, include_top=False, weights='imagenet')
resnet_base = tf.keras.applications.ResNet101V2(input_shape=IMG_SHAPE, include_top=False, weights='imagenet')

NMZivkovic

class Wrapper(tf.keras.Model):
    def __init__(self, base_model):
        super(Wrapper, self).__init__()
        
        self.base_model = base_model
        self.average_pooling_layer = tf.keras.layers.GlobalAveragePooling2D()
        self.output_layer = tf.keras.layers.Dense(1)
        
    def call(self, inputs):
        x = self.base_model(inputs)

NMZivkovic

vgg16_base = tf.keras.applications.VGG16(input_shape=IMG_SHAPE, include_top=False, weights='imagenet')
googlenet_base = tf.keras.applications.InceptionV3(input_shape=IMG_SHAPE, include_top=False, weights='imagenet')
resnet_base = tf.keras.applications.ResNet101V2(input_shape=IMG_SHAPE, include_top=False, weights='imagenet')

NMZivkovic

import numpy as np
import matplotlib.pyplot as plt

import tensorflow as tf
import tensorflow_datasets as tfds

IMG_SIZE = 160
BATCH_SIZE = 32
SHUFFLE_SIZE = 1000
IMG_SHAPE = (IMG_SIZE, IMG_SIZE, 3)

NMZivkovic

data_loader = DataLoader(IMG_SIZE, BATCH_SIZE)

plt.figure(figsize=(10, 8))
i = 0
for img, label in data_loader.get_random_raw_images(20):
    plt.subplot(4, 5, i+1)
    plt.imshow(img)
    plt.title("{} - {}".format(data_loader.get_label_name(label), img.shape))
    plt.xticks([])
    plt.yticks([])

NMZivkovic

def __init__(self, image_size, batch_size):

    self.image_size = image_size
    self.batch_size = batch_size

    # 80% train data, 10% validation data, 10% test data
    split_weights = (8, 1, 1)
    splits = tfds.Split.TRAIN.subsplit(weighted=split_weights)

    (self.train_data_raw, self.validation_data_raw, self.test_data_raw), self.metadata = tfds.load(