Définir les utilitaires Resizer apprenables
def conv_block(x, filters, kernel_size, strides, activation=layers.LeakyReLU(0.2)):
x = layers.Conv2D(filters, kernel_size, strides, padding="same", use_bias=False)(x)
x = layers.BatchNormalization()(x)
if activation:
x = activation(x)
return x
def res_block(x):
inputs = x
x = conv_block(x, 16, 3, 1)
x = conv_block(x, 16, 3, 1, activation=None)
return layers.Add()([inputs, x])
def get_learnable_resizer(filters=16, num_res_blocks=1, interpolation=INTERPOLATION):
inputs = layers.Input(shape=[None, None, 3])
# First, perform naive resizing.
naive_resize = layers.Resizing(
*TARGET_SIZE, interpolation=interpolation
)(inputs)
# First convolution block without batch normalization.
x = layers.Conv2D(filters=filters, kernel_size=7, strides=1, padding="same")(inputs)
x = layers.LeakyReLU(0.2)(x)
# Second convolution block with batch normalization.
x = layers.Conv2D(filters=filters, kernel_size=1, strides=1, padding="same")(x)
x = layers.LeakyReLU(0.2)(x)
x = layers.BatchNormalization()(x)
# Intermediate resizing as a bottleneck.
bottleneck = layers.Resizing(
*TARGET_SIZE, interpolation=interpolation
)(x)
# Residual passes.
for _ in range(num_res_blocks):
x = res_block(bottleneck)
# Projection.
x = layers.Conv2D(
filters=filters, kernel_size=3, strides=1, padding="same", use_bias=False
)(x)
x = layers.BatchNormalization()(x)
# Skip connection.
x = layers.Add()([bottleneck, x])
# Final resized image.
x = layers.Conv2D(filters=3, kernel_size=7, strides=1, padding="same")(x)
final_resize = layers.Add()([naive_resize, x])
return tf.keras.Model(inputs, final_resize, name="learnable_resizer")
learnable_resizer = get_learnable_resizer()
Akshay R