码迷,mamicode.com
首页 > Web开发 > 详细

CS231n assignment3 Q5 Generative Adversarial Networks

时间:2019-01-05 15:46:33      阅读:649      评论:0      收藏:0      [点我收藏+]

标签:inpu   def   HERE   Dimension   max   and   vat   pool   ann   

LeakyReLU

def leaky_relu(x, alpha=0.01):
    """Compute the leaky ReLU activation function.
    
    Inputs:
    - x: TensorFlow Tensor with arbitrary shape
    - alpha: leak parameter for leaky ReLU
    
    Returns:
    TensorFlow Tensor with the same shape as x
    """
    # TODO: implement leaky ReLU
    condition = tf.less(x,0)
    res = tf.where(condition,alpha * x,x)
    return res

Random Noise

def sample_noise(batch_size, dim):
    """Generate random uniform noise from -1 to 1.
    
    Inputs:
    - batch_size: integer giving the batch size of noise to generate
    - dim: integer giving the dimension of the the noise to generate
    
    Returns:
    TensorFlow Tensor containing uniform noise in [-1, 1] with shape [batch_size, dim]
    """
    # TODO: sample and return noise
    return tf.random_uniform([batch_size,dim],minval = -1,maxval = 1)

Discriminator

Architecture:

  • Fully connected layer with input size 784 and output size 256
  • LeakyReLU with alpha 0.01
  • Fully connected layer with output size 256
  • LeakyReLU with alpha 0.01
  • Fully connected layer with output size 1
def discriminator(x):
    """Compute discriminator score for a batch of input images.
    
    Inputs:
    - x: TensorFlow Tensor of flattened input images, shape [batch_size, 784]
    
    Returns:
    TensorFlow Tensor with shape [batch_size, 1], containing the score 
    for an image being real for each input image.
    """
    with tf.variable_scope("discriminator"):
        # TODO: implement architecture
        fc1 = tf.layers.dense(x,256,use_bias = True,name = ‘fc1‘)
        leaky_relu1 = leaky_relu(fc1,alpha = 0.01)
        fc2 = tf.layers.dense(leaky_relu1,256,use_bias = True,name = ‘fc2‘)
        leaky_relu2 = leaky_relu(fc2,alpha = 0.01)
        logits = tf.layers.dense(leaky_relu2,1,name = ‘fc3‘)
        return logits

Generator

Architecture:

  • Fully connected layer with inupt size tf.shape(z)[1] (the number of noise dimensions) and output size 1024
  • ReLU
  • Fully connected layer with output size 1024
  • ReLU
  • Fully connected layer with output size 784
  • TanH (To restrict every element of the output to be in the range [-1,1])
def generator(z):
    """Generate images from a random noise vector.
    
    Inputs:
    - z: TensorFlow Tensor of random noise with shape [batch_size, noise_dim]
    
    Returns:
    TensorFlow Tensor of generated images, with shape [batch_size, 784].
    """
    with tf.variable_scope("generator"):
        # TODO: implement architecture
        fc1 = tf.layers.dense(z,1024,use_bias = True,activation = tf.nn.relu)
        fc2 = tf.layers.dense(fc1,1024,use_bias = True,activation = tf.nn.relu)
        img = tf.layers.dense(fc2,784,use_bias = True,activation = tf.nn.tanh)
        return img

Gan loss

def gan_loss(logits_real, logits_fake):
    """Compute the GAN loss.
    
    Inputs:
    - logits_real: Tensor, shape [batch_size, 1], output of discriminator
        Unnormalized score that the image is real for each real image
    - logits_fake: Tensor, shape[batch_size, 1], output of discriminator
        Unnormalized score that the image is real for each fake image
    
    Returns:
    - D_loss: discriminator loss scalar
    - G_loss: generator loss scalar
    
    HINT: for the discriminator loss, you‘ll want to do the averaging separately for
    its two components, and then add them together (instead of averaging once at the very end).
    """
    # TODO: compute D_loss and G_loss
    loss1 = tf.nn.sigmoid_cross_entropy_with_logits(labels = tf.ones_like(logits_real),logits = logits_real,name = ‘discriminator_real_loss‘)
    loss2 = tf.nn.sigmoid_cross_entropy_with_logits(labels = tf.zeros_like(logits_fake),logits = logits_fake,name = ‘discriminator_fake_loss‘)
    loss3 = tf.nn.sigmoid_cross_entropy_with_logits(labels = tf.ones_like(logits_fake),logits = logits_fake,name = ‘generator_loss‘)
    D_loss = tf.reduce_mean(loss1 + loss2)
    G_loss = tf.reduce_mean(loss3)
    return D_loss, G_loss

Optimizing

# TODO: create an AdamOptimizer for D_solver and G_solver
def get_solvers(learning_rate=1e-3, beta1=0.5):
    """Create solvers for GAN training.
    
    Inputs:
    - learning_rate: learning rate to use for both solvers
    - beta1: beta1 parameter for both solvers (first moment decay)
    
    Returns:
    - D_solver: instance of tf.train.AdamOptimizer with correct learning_rate and beta1
    - G_solver: instance of tf.train.AdamOptimizer with correct learning_rate and beta1
    """
    D_solver = tf.train.AdamOptimizer(learning_rate = learning_rate,beta1 = beta1)
    G_solver = tf.train.AdamOptimizer(learning_rate = learning_rate,beta1 = beta1)
    return D_solver, G_solver

Least Squares GAN

def lsgan_loss(scores_real, scores_fake):
    """Compute the Least Squares GAN loss.
    
    Inputs:
    - scores_real: Tensor, shape [batch_size, 1], output of discriminator
        The score for each real image
    - scores_fake: Tensor, shape[batch_size, 1], output of discriminator
        The score for each fake image    
          
    Returns:
    - D_loss: discriminator loss scalar
    - G_loss: generator loss scalar
    """
    # TODO: compute D_loss and G_loss
    D_loss = 0.5 * tf.reduce_mean(tf.square(scores_real - 1)) + 0.5 * tf.reduce_mean(tf.square(scores_fake))
    G_loss = 0.5 * tf.reduce_mean(tf.square(scores_fake - 1))
    return D_loss, G_loss

Deep Convolutional GANs

Discriminator

Architecture:

  • Conv2D: 32 Filters, 5x5, Stride 1, padding 0
  • Leaky ReLU(alpha=0.01)
  • Max Pool 2x2, Stride 2
  • Conv2D: 64 Filters, 5x5, Stride 1, padding 0
  • Leaky ReLU(alpha=0.01)
  • Max Pool 2x2, Stride 2
  • Flatten
  • Fully Connected with output size 4 x 4 x 64
  • Leaky ReLU(alpha=0.01)
  • Fully Connected with output size 1
def discriminator(x):
    """Compute discriminator score for a batch of input images.
    
    Inputs:
    - x: TensorFlow Tensor of flattened input images, shape [batch_size, 784]
    
    Returns:
    TensorFlow Tensor with shape [batch_size, 1], containing the score 
    for an image being real for each input image.
    """
    x = tf.reshape(x,shape = (tf.shape(x)[0],28,28,1))
    with tf.variable_scope("discriminator"):
        # TODO: implement architecture
        conv1 = tf.layers.conv2d(x,filters = 32,kernel_size = (5,5),strides = (1,1),activation = leaky_relu)
        max_pool1 = tf.layers.max_pooling2d(conv1,pool_size = (2,2),strides = (2,2))
        conv2 = tf.layers.conv2d(max_pool1,filters = 64,kernel_size = (5,5),strides = (1,1),activation = leaky_relu)
        max_pool2 = tf.layers.max_pooling2d(conv2,pool_size = (2,2),strides = (2,2))
        flat = tf.contrib.layers.flatten(max_pool2)
        fc1 = tf.layers.dense(flat,4*4*64,activation = leaky_relu)
        logits = tf.layers.dense(fc1,1)
        return logits
test_discriminator(1102721)

Generator

Architecture:

  • Fully connected with output size 1024
  • ReLU
  • BatchNorm
  • Fully connected with output size 7 x 7 x 128
  • ReLU
  • BatchNorm
  • Resize into Image Tensor of size 7, 7, 128
  • Conv2D^T (transpose): 64 filters of 4x4, stride 2
  • ReLU
  • BatchNorm
  • Conv2d^T (transpose): 1 filter of 4x4, stride 2
  • TanH
def generator(z):
    """Generate images from a random noise vector.
    
    Inputs:
    - z: TensorFlow Tensor of random noise with shape [batch_size, noise_dim]
    
    Returns:
    TensorFlow Tensor of generated images, with shape [batch_size, 784].
    """
    batch_size = tf.shape(z)[0]
    with tf.variable_scope("generator"):
        # TODO: implement architecture
        fc1 = tf.layers.dense(z,1024,activation = tf.nn.relu,use_bias = True)
        bn1 = tf.layers.batch_normalization(fc1,training = True)
        fc2 = tf.layers.dense(bn1,7*7*128,activation = tf.nn.relu,use_bias = True)
        bn2 = tf.layers.batch_normalization(fc2,training = True)
        resize = tf.reshape(bn2,shape = (-1,7,7,128))
        filter_conv1 = tf.get_variable(‘deconv1‘,[4,4,64,128]) # [height, width, output_channels, in_channels]
        conv_tr1 = tf.nn.conv2d_transpose(resize,filter = filter_conv1,output_shape = [batch_size,14,14,64],strides = [1,2,2,1])
        bias1 = tf.get_variable(‘deconv1_bias‘,[64])
        conv_tr1 += bias1
        relu_conv_tr1 = tf.nn.relu(conv_tr1)
        bn3 = tf.layers.batch_normalization(relu_conv_tr1,training = True)
        filter_conv2 = tf.get_variable(‘deconv2‘,[4,4,1,64])
        conv_tr2 = tf.nn.conv2d_transpose(bn3,filter = filter_conv2,output_shape = [batch_size,28,28,1],strides = [1,2,2,1])
        bias2 = tf.get_variable(‘deconv2_bias‘,[1])
        conv_tr2 += bias2
        img = tf.nn.tanh(conv_tr2)
        img = tf.contrib.layers.flatten(img)
        return img
test_generator(6595521)

技术分享图片

Epoch: 0, D: 0.2112, G:0.3559
Epoch: 1, D: 0.196, G:0.2681

技术分享图片

Epoch: 2, D: 0.1689, G:0.2728
Epoch: 3, D: 0.1618, G:0.2215

技术分享图片

Epoch: 4, D: 0.1968, G:0.2461
Epoch: 5, D: 0.1968, G:0.2429

技术分享图片

Epoch: 6, D: 0.2316, G:0.1997
Epoch: 7, D: 0.2206, G:0.1858

技术分享图片

Epoch: 8, D: 0.2131, G:0.1815
Epoch: 9, D: 0.2345, G:0.1732
Final images

技术分享图片

CS231n assignment3 Q5 Generative Adversarial Networks

标签:inpu   def   HERE   Dimension   max   and   vat   pool   ann   

原文地址:https://www.cnblogs.com/bernieloveslife/p/10224687.html

(0)
(0)
   
举报
评论 一句话评论(0
登录后才能评论!
© 2014 mamicode.com 版权所有  联系我们:gaon5@hotmail.com
迷上了代码!