Implementation of ACGAN Model
Imports
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
import numpy as npmodels = tf.contrib.keras.models
layers = tf.contrib.keras.layers
utils = tf.contrib.keras.utils
losses = tf.contrib.keras.losses
optimizers = tf.contrib.keras.optimizers
metrics = tf.contrib.keras.metricsConstruct Generator
def generator(latent_size, classes=10):
def up_sampling_block(x, filter_size):
# upsample block
x = layers.UpSampling2D(size=(2,2))(x)
x = layers.Conv2D(filter_size, (5,5), padding='same', activation='relu')(x)
return x
# Input 1
# image class label
image_class = layers.Input(shape=(1,), dtype='int32', name='image_class')
# class embeddings
emb = layers.Embedding(classes, latent_size,
embeddings_initializer='glorot_normal')(image_class)
# 10 classes in MNIST
cls = layers.Flatten()(emb)
# Input 2
# latent noise vector
latent_input = layers.Input(shape=(latent_size,), name='latent_noise')
# hadamard product between latent embedding and a class conditional embedding
h = layers.multiply([latent_input, cls])
# Conv generator
x = layers.Dense(1024, activation='relu')(h)
x = layers.Dense(128 * 7 * 7, activation='relu')(x)
x = layers.Reshape((7, 7, 128))(x)
# upsample to (14, 14, 128)
x = up_sampling_block(x, 128)
# upsample to (28, 28, 256)
x = up_sampling_block(x, 256)
# reduce channel into binary image (28, 28, 1)
generated_img = layers.Conv2D(1, (2,2), padding='same', activation='tanh')(x)
return models.Model(inputs=[latent_input, image_class],
outputs=generated_img,
name='generator')g = generator(latent_size = 100, classes=10)
g.summary()Layer (type) Output Shape Param # Connected to
====================================================================================================
image_class (InputLayer) (None, 1) 0
____________________________________________________________________________________________________
embedding_6 (Embedding) (None, 1, 100) 1000 image_class[0][0]
____________________________________________________________________________________________________
latent_noise (InputLayer) (None, 100) 0
____________________________________________________________________________________________________
flatten_10 (Flatten) (None, 100) 0 embedding_6[0][0]
____________________________________________________________________________________________________
multiply_6 (Multiply) (None, 100) 0 latent_noise[0][0]
flatten_10[0][0]
____________________________________________________________________________________________________
dense_11 (Dense) (None, 1024) 103424 multiply_6[0][0]
____________________________________________________________________________________________________
dense_12 (Dense) (None, 6272) 6428800 dense_11[0][0]
____________________________________________________________________________________________________
reshape_6 (Reshape) (None, 7, 7, 128) 0 dense_12[0][0]
____________________________________________________________________________________________________
up_sampling2d_11 (UpSampling2D) (None, 14, 14, 128) 0 reshape_6[0][0]
____________________________________________________________________________________________________
conv2d_32 (Conv2D) (None, 14, 14, 256) 819456 up_sampling2d_11[0][0]
____________________________________________________________________________________________________
up_sampling2d_12 (UpSampling2D) (None, 28, 28, 256) 0 conv2d_32[0][0]
____________________________________________________________________________________________________
conv2d_33 (Conv2D) (None, 28, 28, 128) 819328 up_sampling2d_12[0][0]
____________________________________________________________________________________________________
conv2d_34 (Conv2D) (None, 28, 28, 1) 513 conv2d_33[0][0]
====================================================================================================
Total params: 8,172,521
Trainable params: 8,172,521
Non-trainable params: 0
____________________________________________________________________________________________________Construct Discriminator
def discriminator(input_shape=(28, 28, 1)):
def conv_block(x, filter_size, stride):
x = layers.Conv2D(filter_size, (3,3), padding='same', strides=stride)(x)
x = layers.LeakyReLU()(x)
x = layers.Dropout(0.3)(x)
return x
input_img = layers.Input(shape=input_shape)
# discriminator network
x = conv_block(input_img, 32, (2,2))
x = conv_block(input_img, 64, (1,1))
x = conv_block(input_img, 128, (2,2))
x = conv_block(input_img, 256, (1,1))
features = layers.Flatten()(x)
# binary classifier, image fake or real
fake = layers.Dense(1, activation='sigmoid', name='generation')(features)
# multi-class classifier, image digit class
aux = layers.Dense(10, activation='softmax', name='auxiliary')(features)
return models.Model(inputs=input_img, outputs=[fake, aux], name='discriminator')d = discriminator(input_shape=(28, 28, 1))
d.summary()Layer (type) Output Shape Param # Connected to
====================================================================================================
input_6 (InputLayer) (None, 28, 28, 1) 0
____________________________________________________________________________________________________
conv2d_35 (Conv2D) (None, 14, 14, 32) 320 input_6[0][0]
____________________________________________________________________________________________________
leaky_re_lu_17 (LeakyReLU) (None, 14, 14, 32) 0 conv2d_35[0][0]
____________________________________________________________________________________________________
dropout_17 (Dropout) (None, 14, 14, 32) 0 leaky_re_lu_17[0][0]
____________________________________________________________________________________________________
conv2d_36 (Conv2D) (None, 14, 14, 64) 18496 dropout_17[0][0]
____________________________________________________________________________________________________
leaky_re_lu_18 (LeakyReLU) (None, 14, 14, 64) 0 conv2d_36[0][0]
____________________________________________________________________________________________________
dropout_18 (Dropout) (None, 14, 14, 64) 0 leaky_re_lu_18[0][0]
____________________________________________________________________________________________________
conv2d_37 (Conv2D) (None, 7, 7, 128) 73856 dropout_18[0][0]
____________________________________________________________________________________________________
leaky_re_lu_19 (LeakyReLU) (None, 7, 7, 128) 0 conv2d_37[0][0]
____________________________________________________________________________________________________
dropout_19 (Dropout) (None, 7, 7, 128) 0 leaky_re_lu_19[0][0]
____________________________________________________________________________________________________
conv2d_38 (Conv2D) (None, 7, 7, 256) 295168 dropout_19[0][0]
____________________________________________________________________________________________________
leaky_re_lu_20 (LeakyReLU) (None, 7, 7, 256) 0 conv2d_38[0][0]
____________________________________________________________________________________________________
dropout_20 (Dropout) (None, 7, 7, 256) 0 leaky_re_lu_20[0][0]
____________________________________________________________________________________________________
flatten_11 (Flatten) (None, 12544) 0 dropout_20[0][0]
____________________________________________________________________________________________________
generation (Dense) (None, 1) 12545 flatten_11[0][0]
____________________________________________________________________________________________________
auxiliary (Dense) (None, 10) 125450 flatten_11[0][0]
====================================================================================================
Total params: 525,835
Trainable params: 525,835
Non-trainable params: 0
____________________________________________________________________________________________________Combine Generator with Discriminator
# Adam parameters suggested in paper
adam_lr = 0.0002
adam_beta_1 = 0.5
def ACGAN(latent_size = 100):
# build the discriminator
dis = discriminator()
dis.compile(
optimizer=optimizers.Adam(lr=adam_lr, beta_1=adam_beta_1),
loss=['binary_crossentropy', 'sparse_categorical_crossentropy']
)
# build the generator
gen = generator(latent_size)
gen.compile(optimizer=optimizers.Adam(lr=adam_lr, beta_1=adam_beta_1),
loss='binary_crossentropy')
# Inputs
latent = layers.Input(shape=(latent_size, ), name='latent_noise')
image_class = layers.Input(shape=(1,), dtype='int32', name='image_class')
# Get a fake image
fake_img = gen([latent, image_class])
# Only train generator in combined model
dis.trainable = False
fake, aux = dis(fake_img)
combined = models.Model(inputs=[latent, image_class],
outputs=[fake, aux],
name='ACGAN')
combined.compile(
optimizer=optimizers.Adam(lr=adam_lr, beta_1=adam_beta_1),
loss=['binary_crossentropy', 'sparse_categorical_crossentropy']
)
return combinedcagan = ACGAN(latent_size = 100)
cagan.summary()Layer (type) Output Shape Param # Connected to
====================================================================================================
latent_noise (InputLayer) (None, 100) 0
____________________________________________________________________________________________________
image_class (InputLayer) (None, 1) 0
____________________________________________________________________________________________________
generator (Model) (None, 28, 28, 1) 8172521 latent_noise[0][0]
image_class[0][0]
____________________________________________________________________________________________________
discriminator (Model) [(None, 1), (None, 10 525835 generator[1][0]
====================================================================================================
Total params: 8,698,356
Trainable params: 8,172,521
Non-trainable params: 525,835
____________________________________________________________________________________________________Next Lesson
Train and Evaluate ACGAN
Image classification task with MNIST
Last updated
