doleron

output = Lambda(euclidean_distance, name="output_layer")([vect_output_a, vect_output_b])

doleron

INPUT_SHAPE = (INPUT_SIZE, INPUT_SIZE, 3)

def initialize_base_network():

    inputs = tf.keras.layers.Input(INPUT_SHAPE)
    base_model = tf.keras.applications.mobilenet_v2.MobileNetV2(input_shape=INPUT_SHAPE, include_top=False, weights='imagenet')
    
    base_model.trainable = True
    fine_tune_at = len(base_model.layers)-int(len(base_model.layers)*.10)
    for layer in base_model.layers[:fine_tune_at]:

doleron

def buid_model():
    
    base_network = initialize_base_network()

    input_a = Input(shape=INPUT_SHAPE, name="left_input")
    vect_output_a = base_network(input_a)

    input_b = Input(shape=INPUT_SHAPE, name="right_input")
    vect_output_b = base_network(input_b)

doleron

VALIDATION_BATCH_SIZE = 2

def build_validation_dataset():

    pairs_tensor = tf.convert_to_tensor(validation_pairs)
    labels_tensor = tf.convert_to_tensor(validation_pairs_labels)

    result = tf.data.Dataset.from_tensor_slices((pairs_tensor, labels_tensor))

    result = result.map(lambda pair, label: (load_images(pair), label))

doleron

INPUT_SIZE = 244
TRAINING_BATCH_SIZE = 8

def load_image(file_name):
    raw = tf.io.read_file(file_name)
    image = tf.io.decode_image(raw, expand_animations = False, channels=3)
    image = tf.image.resize(image, size=(INPUT_SIZE, INPUT_SIZE), preserve_aspect_ratio=True)
    image = tf.image.resize_with_crop_or_pad(image, INPUT_SIZE, INPUT_SIZE)
    image = tf.cast(image, tf.float32) / 255.0
    return image

doleron

def build_training_dataset():

    pairs_tensor = tf.convert_to_tensor(training_pairs)
    labels_tensor = tf.convert_to_tensor(training_pairs_labels)

    result = tf.data.Dataset.from_tensor_slices((pairs_tensor, labels_tensor))

    result = result.map(lambda pair, label: (load_images(pair), label))
    result = result.shuffle(100, reshuffle_each_iteration=True)
    result = result.repeat()

doleron

early_stop = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience = 10)

history = model.fit(train_ds, validation_data=validation_ds, epochs=EPOCHS, callbacks=[early_stop])

doleron

data_augmentation = tf.keras.Sequential([
    tf.keras.layers.RandomFlip("horizontal"),
    tf.keras.layers.RandomRotation(0.01),
    tf.keras.layers.RandomBrightness(factor=0.2, value_range=(0., 1.)),
    tf.keras.layers.GaussianNoise(0.002),
    tf.keras.layers.RandomZoom(height_factor=(-0.1, 0.1))
])

doleron

from tensorflow.keras import backend as K

def euclidean_distance(x, y):
    sum_square = K.sum(K.square(x - y), axis=1, keepdims=True)
    return K.sqrt(K.maximum(sum_square, K.epsilon()))

doleron

plt.figure(figsize=(12, 10))

test_list = list(test_ds.take(20).as_numpy_iterator())

image, labels = test_list[0]

for i in range(len(test_list)):

    ax = plt.subplot(4, 5, i + 1)
    image, labels = test_list[i]