DCGAN

The implementation of Deep Convolutional Generative Adversarial Network (DCGAN)

This is a basic implementation of Deep Convolutional Generative Adversarial Network (DCGAN). The code mainly refers to the TensorFlow Tutorials in tensorflow/docs. Two datasets mnist and Large-scale Celeb Faces Attributes (CelebA) were used.

Requirements

• python3
• tensorflow-gpu == 2.0.0

The CelebA dataset can be downloaded in kaggle by the above link.

Getting start

• Download the dataset CelebA in kaggle
• Modify the dataset path in dcgan.py
• Run the script dcgan.py (ensure you have at least one GPU device) and image_at_epoch_xxxx.png will be generated in result folder
• Run the script generate_visual_result.py and ground_true_image.png, dcgan.gif, final.png will be generated.
• Enjoy it!

When apply to the mnist dataset, the first and second step can be ignored.

There are some differences in the newotk architecture that implements these two data sets

The detail can be seem in the code.

• Generator

  • mnist

    #The Generator
    #Input a seed (random noise). Start with a Dense layer, then upsample by ConvTranspose.
    #input shape = batch_size * 100
    #output shape = 28 * 28 * 1
    def make_generator_model():
        model = tf.keras.Sequential()
        model.add(layers.Dense(7*7*256, use_bias=False, input_shape=(100, )))
        model.add(layers.BatchNormalization())
        model.add(layers.LeakyReLU())
    
        model.add(layers.Reshape((7, 7, 256))) # (bs, 7, 7, 256)
    
        model.add(layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False)) # (bs, 7, 7, 128)
        model.add(layers.BatchNormalization())
        model.add(layers.LeakyReLU())
    
        model.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False)) # (bs, 14, 14, 64)
        model.add(layers.BatchNormalization())
        model.add(layers.LeakyReLU())
    
        model.add(layers.Conv2DTranspose(1, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh')) # (bs, 28, 28, 1)
    
        return model

  • CelebA

    #The Generator
    #Input a seed (random noise). Start with a Dense layer, then upsample by ConvTranspose.
    #input shape = batch_size * 500
    #output shape = 128 * 128 * 3
    def make_generator_model():
        model = tf.keras.Sequential()
        model.add(layers.Dense(16*16*256, use_bias=False, input_shape=(500, )))
        model.add(layers.BatchNormalization())
        model.add(layers.LeakyReLU())
    
        model.add(layers.Reshape((16, 16, 256))) # (bs, 16, 16, 256)
    
        model.add(layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False)) # (bs, 16, 16, 128)
        model.add(layers.BatchNormalization())
        model.add(layers.LeakyReLU())
    
        model.add(layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False)) # (bs, 32, 32, 64)
        model.add(layers.BatchNormalization())
        model.add(layers.LeakyReLU())
    
        model.add(layers.Conv2DTranspose(32, (5, 5), strides=(2, 2), padding='same', use_bias=False)) # (bs, 64, 64, 32)
        model.add(layers.BatchNormalization())
        model.add(layers.LeakyReLU())
    
        model.add(layers.Conv2DTranspose(3, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh')) # (bs, 128, 128, 3)
    
        return model

• Discriminator

  • mnist

    #The Discriminator
    #A CNN-based image classifier
    #input shape = batch_size * 28 * 28 * 1
    #output shape = batch_size * 1
    def make_discriminator_model():
        model = tf.keras.Sequential()
        model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same', input_shape=[28, 28, 1])) # (bs, 14, 14, 64)
        model.add(layers.LeakyReLU())
        model.add(layers.Dropout(0.3))
    
        model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same')) # (bs, 7, 7, 128)
        model.add(layers.LeakyReLU())
        model.add(layers.Dropout(0.3))
    
        model.add(layers.Flatten()) # (bs, 7*7*128)
        model.add(layers.Dense(1))
    
        return model

  • CelebA

    #The Discriminator
    #A CNN-based image classifier
    #input shape = batch_size * 128 * 128 * 3
    #output shape = batch_size * 1
    def make_discriminator_model():
        model = tf.keras.Sequential()
        model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same', input_shape=[128, 128, 3])) # (bs, 64, 64, 64)
        model.add(layers.LeakyReLU())
        model.add(layers.Dropout(0.3))
    
        model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same')) # (bs, 32, 32, 128)
        model.add(layers.LeakyReLU())
        model.add(layers.Dropout(0.3))
    
        model.add(layers.Conv2D(256, (5, 5), strides=(2, 2), padding='same')) # (bs, 16, 16, 256)
        model.add(layers.LeakyReLU())
        model.add(layers.Dropout(0.3))
    
        model.add(layers.Flatten()) # (bs, 16*16*256)
        model.add(layers.Dense(1))
    
        return model

Visual result

mnist

ground true

generating process

final generated result

CelebA

ground true

generating process

final generated result

LICENSE

The code is under Apache-2.0 License.