DCGAN Tensorflow 代码不会在 CelebA 数据集上生成人脸
DCGAN Tensorflow code doesn't produce faces on CelebA dataset
我已经编写了以下代码，但没有在 celebA 数据集上生成人脸。我认为它应该在每个纪元的最后一次迭代中创建某种面孔（即使非常模糊）。但是，它只会创建没有可见面孔的嘈杂方块。我对 GAN 很陌生，我不确定如何调试这个深度卷积 GAN (DCGAN) 以找出问题所在。
我的代码在这里可能更容易被看到：
https://pastebin.com/c4QUqxJy
这是代码：
from __future__ import print_function
import random
import os
import glob
import scipy
import tensorflow as tf
import numpy as np
from PIL import Image
import skimage.io as io
import matplotlib.pyplot as plt
class Arguments(object):
data_path = 'results_celebA/preprocessed/'
save_path = 'results_celebA' #path to save preprocessed image folder
preproc_foldername = 'preprocessed' #folder name for preprocessed images
image_size = 64 #images are resized to image_size value
num_images = 202590 #the number of training images
batch_size = 64 #batch size
dim_z = 100 #the dimension of z variable (the generator input dimension)
n_g_filters = 64 #the number of the generator filters (gets multiplied between layers)
n_f_filters = 64 #the number of the discriminator filters (gets multiplied between layers)
n_epoch = 25 #the number of epochs
lr = 0.0002 #learning rate
beta1 = 0.5 #beta_1 parameter of Adam optimizer
beta2 = 0.99 #beta_2 parameter of Adam optimizer
args = Arguments()
#contains functions that load, preprocess and visualize images.
class Dataset(object):
def __init__(self, data_path, num_imgs, target_imgsize):
self.data_path = data_path
self.num_imgs = num_imgs
self.target_imgsize = target_imgsize
def normalize_np_image(self, image):
return (image / 255.0 - 0.5) / 0.5
def denormalize_np_image(self, image):
return (image * 0.5 + 0.5) * 255
def get_input(self, image_path):
image = np.array(Image.open(image_path)).astype(np.float32)
return self.normalize_np_image(image)
def get_imagelist(self, data_path, celebA=False):
if celebA == True:
imgs_path = os.path.join(data_path, 'img_align_celeba/*.jpg')
else:
imgs_path = os.path.join(data_path, '*.jpg')
all_namelist = glob.glob(imgs_path, recursive=True)
return all_namelist[:self.num_imgs]
def load_and_preprocess_image(self, image_path):
image = Image.open(image_path)
j = (image.size[0] - 100) // 2
i = (image.size[1] - 100) // 2
image = image.crop([j, i, j + 100, i + 100])
image = image.resize([self.target_imgsize, self.target_imgsize], Image.BILINEAR)
image = np.array(image.convert('RGB')).astype(np.float32)
image = self.normalize_np_image(image)
return image
#reads data, preprocesses and saves to another folder with the given path.
def preprocess_and_save_images(self, dir_name, save_path=''):
preproc_folder_path = os.path.join(save_path, dir_name)
if not os.path.exists(preproc_folder_path):
os.makedirs(preproc_folder_path)
imgs_path = os.path.join(self.data_path, 'img_align_celeba/*.jpg')
print('Saving and preprocessing images ...')
for num, imgname in enumerate(glob.iglob(imgs_path, recursive=True)):
cur_image = self.load_and_preprocess_image(imgname)
cur_image = Image.fromarray(np.uint8(self.denormalize_np_image(cur_image)))
cur_image.save(preproc_folder_path + '/preprocessed_image_%d.jpg' %(num))
self.data_path= preproc_folder_path
def get_nextbatch(self, batch_size):
print("nextbatch batchsize is: ", batch_size)
assert (batch_size > 0),"Give a valid batch size"
cur_idx = 0
image_namelist = self.get_imagelist(self.data_path)
while cur_idx + batch_size <= self.num_imgs:
cur_namelist = image_namelist[cur_idx:cur_idx + batch_size]
cur_batch = [self.get_input(image_path) for image_path in cur_namelist]
cur_batch = np.array(cur_batch).astype(np.float32)
cur_idx += batch_size
yield cur_batch
def show_image(self, image, normalized=True):
if not type(image).__module__ == np.__name__:
image = image.numpy()
if normalized:
npimg = (image * 0.5) + 0.5
npimg.astype(np.uint8)
plt.imshow(npimg, interpolation='nearest')
#contains functions that load, preprocess and visualize images.
class Dataset(object):
def __init__(self, data_path, num_imgs, target_imgsize):
self.data_path = data_path
self.num_imgs = num_imgs
self.target_imgsize = target_imgsize
def normalize_np_image(self, image):
return (image / 255.0 - 0.5) / 0.5
def denormalize_np_image(self, image):
return (image * 0.5 + 0.5) * 255
def get_input(self, image_path):
image = np.array(Image.open(image_path)).astype(np.float32)
return self.normalize_np_image(image)
def get_imagelist(self, data_path, celebA=False):
if celebA == True:
imgs_path = os.path.join(data_path, 'img_align_celeba/*.jpg')
else:
imgs_path = os.path.join(data_path, '*.jpg')
all_namelist = glob.glob(imgs_path, recursive=True)
return all_namelist[:self.num_imgs]
def load_and_preprocess_image(self, image_path):
image = Image.open(image_path)
j = (image.size[0] - 100) // 2
i = (image.size[1] - 100) // 2
image = image.crop([j, i, j + 100, i + 100])
image = image.resize([self.target_imgsize, self.target_imgsize], Image.BILINEAR)
image = np.array(image.convert('RGB')).astype(np.float32)
image = self.normalize_np_image(image)
return image
#reads data, preprocesses and saves to another folder with the given path.
def preprocess_and_save_images(self, dir_name, save_path=''):
preproc_folder_path = os.path.join(save_path, dir_name)
if not os.path.exists(preproc_folder_path):
os.makedirs(preproc_folder_path)
imgs_path = os.path.join(self.data_path, 'img_align_celeba/*.jpg')
print('Saving and preprocessing images ...')
for num, imgname in enumerate(glob.iglob(imgs_path, recursive=True)):
cur_image = self.load_and_preprocess_image(imgname)
cur_image = Image.fromarray(np.uint8(self.denormalize_np_image(cur_image)))
cur_image.save(preproc_folder_path + '/preprocessed_image_%d.jpg' %(num))
self.data_path= preproc_folder_path
def get_nextbatch(self, batch_size):
assert (batch_size > 0),"Give a valid batch size"
cur_idx = 0
image_namelist = self.get_imagelist(self.data_path)
while cur_idx + batch_size <= self.num_imgs:
cur_namelist = image_namelist[cur_idx:cur_idx + batch_size]
cur_batch = [self.get_input(image_path) for image_path in cur_namelist]
cur_batch = np.array(cur_batch).astype(np.float32)
cur_idx += batch_size
yield cur_batch
def show_image(self, image, normalized=True):
if not type(image).__module__ == np.__name__:
image = image.numpy()
if normalized:
npimg = (image * 0.5) + 0.5
npimg.astype(np.uint8)
plt.imshow(npimg, interpolation='nearest')
def generator(x, args, reuse=False):
with tf.device('/gpu:0'):
with tf.variable_scope("generator", reuse=reuse):
#Layer Block 1
with tf.variable_scope("layer1"):
deconv1 = tf.layers.conv2d_transpose(inputs=x,
filters= args.n_g_filters*8,
kernel_size=4,
strides=1,
padding='valid',
use_bias=False,
name='deconv')
batch_norm1=tf.layers.batch_normalization(deconv1,
name = 'batch_norm')
relu1 = tf.nn.relu(batch_norm1, name='relu')
#Layer Block 2
with tf.variable_scope("layer2"):
deconv2 = tf.layers.conv2d_transpose(inputs=relu1,
filters=args.n_g_filters*4,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='deconv')
batch_norm2 = tf.layers.batch_normalization(deconv2,
name = 'batch_norm')
relu2 = tf.nn.relu(batch_norm2, name='relu')
#Layer Block 3
with tf.variable_scope("layer3"):
deconv3 = tf.layers.conv2d_transpose(inputs=relu2,
filters=args.n_g_filters*2,
kernel_size=4,
strides=2,
padding='same',
use_bias = False,
name='deconv')
batch_norm3 = tf.layers.batch_normalization(deconv3,
name = 'batch_norm')
relu3 = tf.nn.relu(batch_norm3, name='relu')
#Layer Block 4
with tf.variable_scope("layer4"):
deconv4 = tf.layers.conv2d_transpose(inputs=relu3,
filters=args.n_g_filters,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='deconv')
batch_norm4 = tf.layers.batch_normalization(deconv4,
name = 'batch_norm')
relu4 = tf.nn.relu(batch_norm4, name='relu')
#Output Layer
with tf.variable_scope("last_layer"):
logit = tf.layers.conv2d_transpose(inputs=relu4,
filters=3,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='logit')
output = tf.nn.tanh(logit)
return output, logit
def discriminator(x, args, reuse=False):
with tf.device('/gpu:0'):
with tf.variable_scope("discriminator", reuse=reuse):
with tf.variable_scope("layer1"):
conv1 = tf.layers.conv2d(inputs=x,
filters=args.n_f_filters,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='conv')
relu1 = tf.nn.leaky_relu(conv1, alpha=0.2, name='relu')
with tf.variable_scope("layer2"):
conv2 = tf.layers.conv2d(inputs=relu1,
filters=args.n_f_filters*2,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='conv')
batch_norm2 = tf.layers.batch_normalization(conv2,name='batch_norm')
relu2 = tf.nn.leaky_relu(batch_norm2, alpha=0.2, name='relu')
with tf.variable_scope("layer3"):
conv3 = tf.layers.conv2d(inputs=relu2,
filters=args.n_f_filters*4,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='conv')
batch_norm3 = tf.layers.batch_normalization(conv3, name='batch_norm')
relu3 = tf.nn.leaky_relu(batch_norm3, name='relu')
with tf.variable_scope("layer4"):
conv4 = tf.layers.conv2d(inputs=relu3,
filters=args.n_f_filters*8,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='conv')
batch_norm4 = tf.layers.batch_normalization(conv4, name='batch_norm')
relu4 = tf.nn.leaky_relu(batch_norm4, alpha=0.2, name='relu')
with tf.variable_scope("last_layer"):
logit = tf.layers.conv2d(inputs=relu4,
filters=1,
kernel_size=4,
strides=1,
padding='valid',
use_bias=False,
name='conv')
output = tf.nn.sigmoid(logit)
return output, logit
def sample_z(dim_z, num_batch):
mu = 0
sigma = 1
s = np.random.normal(mu, sigma, num_batch*dim_z)
samples = s.reshape(num_batch, 1, 1, dim_z)
##dist = tf.distributions.Normal(0.0, 1.0)
##samples = dist.sample([num_batch, 1, 1, dim_z])
return samples
#64,1,1,100 6400
sample_z(100, 64)
def get_losses(d_real_logits, d_fake_logits):
#add new loss function here
###d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real_logits, labels=tf.ones_like(d_real_logits)))
###d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits, labels=tf.zeros_like(d_fake_logits)))
###d_loss = d_loss_real + d_loss_fake
###g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits, labels=tf.ones_like(d_fake_logits)))
###return d_loss, g_loss
d_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real_logits,labels=tf.ones_like(d_real_logits)) + tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits,labels=tf.zeros_like(d_fake_logits)))
g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits,labels=tf.ones_like(d_fake_logits)))
return d_loss, g_loss
def get_optimizers(learning_rate, beta1, beta2):
d_optimizer = tf.train.AdamOptimizer(learning_rate, beta1, beta2)
g_optimizer = tf.train.AdamOptimizer(learning_rate, beta1, beta2)
return d_optimizer, g_optimizer
def optimize(d_optimizer, g_optimizer, d_loss, g_loss):
d_step = d_optimizer.minimize(d_loss)
g_step = g_optimizer.minimize(g_loss)
return d_step, g_step
LOGDIR = "logs_basic_dcgan"
def merge_images(image_batch, size):
h,w = image_batch.shape[1], image_batch.shape[2]
c = image_batch.shape[3]
img = np.zeros((int(h*size[0]), w*size[1], c))
for idx, im in enumerate(image_batch):
i = idx % size[1]
j = idx // size[1]
img[j*h:j*h+h, i*w:i*w+w,:] = im
return img
itr_fh = open('basic_gan_itr.txt', 'a+')
def train(args):
tf.reset_default_graph()
data_loader = Dataset(args.data_path, args.num_images, args.image_size)
#data_loader.preprocess_and_save_images('preprocessed', 'results_celebA') #preprocess the images once
X = tf.placeholder(tf.float32, shape=[args.batch_size, args.image_size , args.image_size, 3])
Z = tf.placeholder(tf.float32, shape=[args.batch_size, 1, 1, args.dim_z])
G_sample, _ = generator(Z, args)
D_real, D_real_logits = discriminator(X, args)
D_fake, D_fake_logits = discriminator(G_sample, args, reuse=True)
d_loss, g_loss = get_losses(D_real_logits, D_fake_logits)
d_optimizer, g_optimizer = get_optimizers(args.lr, args.beta1, args.beta2)
d_step, g_step = optimize(d_optimizer, g_optimizer, d_loss, g_loss)
###z_sum = tf.summary.histogram('z', Z)
###d_sum = tf.summary.histogram('d', D_real)
###G_sum = tf.summary.histogram('g', G_sample)
###d_loss_sum = tf.summary.scalar('d_loss', d_loss)
###g_loss_sum = tf.summary.scalar('g_loss', g_loss)
###d_sum = tf.summary.merge([z_sum, d_sum, d_loss_sum])
###g_sum = tf.summary.merge([z_sum, G_sum, g_loss_sum])
###saver = tf.train.Saver()
###merged_summary = tf.summary.merge_all()
###d_loss_summary = tf.summary.scalar("Discriminator_Total_Loss", d_loss)
###g_loss_summary = tf.summary.scalar("Generator_Total_Loss", g_loss)
###merged_summary = tf.summary.merge_all()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(args.n_epoch):
for itr, real_batch in enumerate(data_loader.get_nextbatch(args.batch_size)):
print('itr is %d, and epoch is %d' %(itr, epoch))
itr_fh.write("epoch: " + str(epoch) + " itr: " + str(itr) + "\n")
Z_sample = sample_z(args.dim_z, args.batch_size)
_, _ = sess.run([d_step, g_step], feed_dict={X:real_batch , Z:Z_sample})
sample = sess.run(G_sample, feed_dict={Z:Z_sample})
print("sample size is: ", sample.shape)
if itr==3164: #num_images/batch_size
im_merged = merge_images(sample[:16], [4,4])
plt.imsave('sample_gan_images/im_merged_epoch_%d.png' %(epoch), im_merged )
scipy.misc.imsave('sample_gan_images/im_epoch_%d_itr_%d.png' %(epoch,itr), sample[1])
##merged_summary = sess.run(merged_summary, feed_dict={X:real_batch , Z:Z_sample})
###writer = tf.summary.FileWriter(LOGDIR)
###writer.add_summary(merged_summary, itr)
###d_loss_summary = tf.summary.scalar("Discriminator_Total_Loss", d_loss)
###g_loss_summary = tf.summary.scalar("Generator_Total_Loss", g_loss)
###merged_summary = tf.summary.merge_all()
###writer.add_graph(sess.graph)
###saver.save(sess, save_path='logs_basic_dcgan/gan.ckpt')
train(args)
这是在前 5 个 epoch 结束时创建的图像。我也评论了与 tensorboard 相关的东西，因为不幸的是它使它变得非常慢。
纪元 0 结束：
时代 1 结束：
时代 2 结束：
纪元 3 结束：
时代 4 结束：
我认为问题与优化器的定义有关：
def optimize(d_optimizer, g_optimizer, d_loss, g_loss):
d_step = d_optimizer.minimize(d_loss)
g_step = g_optimizer.minimize(g_loss)
return d_step, g_step
虽然您用相应的损失定义了每个优化器，但您没有传递每个优化器将训练的变量列表。因此，通过 default 函数 minimize 将考虑图形集合 GraphKeys.TRAINABLE_VARIABLES 下的所有变量。由于您的所有变量都在此图形集合下定义，因此当您调用 d_step 和调用 g_step.
时，您当前的代码实际上会更新来自生成器和鉴别器的所有变量
您必须为每个模型定义变量列表。由于您使用的是变量范围，因此一种方法是：
def optimize(d_optimizer, g_optimizer, d_loss, g_loss):
t_vars = tf.trainable_variables()
d_vars = [var for var in t_vars if var.name.startswith('discriminator')]
g_vars = [var for var in t_vars if var.name.startswith('generator')]
d_step = d_optimizer.minimize(d_loss, var_list=d_vars)
g_step = g_optimizer.minimize(g_loss, var_list=g_vars)
return d_step, g_step
我已经编写了以下代码，但没有在 celebA 数据集上生成人脸。我认为它应该在每个纪元的最后一次迭代中创建某种面孔（即使非常模糊）。但是，它只会创建没有可见面孔的嘈杂方块。我对 GAN 很陌生，我不确定如何调试这个深度卷积 GAN (DCGAN) 以找出问题所在。
我的代码在这里可能更容易被看到： https://pastebin.com/c4QUqxJy 这是代码：
from __future__ import print_function
import random
import os
import glob
import scipy
import tensorflow as tf
import numpy as np
from PIL import Image
import skimage.io as io
import matplotlib.pyplot as plt
class Arguments(object):
data_path = 'results_celebA/preprocessed/'
save_path = 'results_celebA' #path to save preprocessed image folder
preproc_foldername = 'preprocessed' #folder name for preprocessed images
image_size = 64 #images are resized to image_size value
num_images = 202590 #the number of training images
batch_size = 64 #batch size
dim_z = 100 #the dimension of z variable (the generator input dimension)
n_g_filters = 64 #the number of the generator filters (gets multiplied between layers)
n_f_filters = 64 #the number of the discriminator filters (gets multiplied between layers)
n_epoch = 25 #the number of epochs
lr = 0.0002 #learning rate
beta1 = 0.5 #beta_1 parameter of Adam optimizer
beta2 = 0.99 #beta_2 parameter of Adam optimizer
args = Arguments()
#contains functions that load, preprocess and visualize images.
class Dataset(object):
def __init__(self, data_path, num_imgs, target_imgsize):
self.data_path = data_path
self.num_imgs = num_imgs
self.target_imgsize = target_imgsize
def normalize_np_image(self, image):
return (image / 255.0 - 0.5) / 0.5
def denormalize_np_image(self, image):
return (image * 0.5 + 0.5) * 255
def get_input(self, image_path):
image = np.array(Image.open(image_path)).astype(np.float32)
return self.normalize_np_image(image)
def get_imagelist(self, data_path, celebA=False):
if celebA == True:
imgs_path = os.path.join(data_path, 'img_align_celeba/*.jpg')
else:
imgs_path = os.path.join(data_path, '*.jpg')
all_namelist = glob.glob(imgs_path, recursive=True)
return all_namelist[:self.num_imgs]
def load_and_preprocess_image(self, image_path):
image = Image.open(image_path)
j = (image.size[0] - 100) // 2
i = (image.size[1] - 100) // 2
image = image.crop([j, i, j + 100, i + 100])
image = image.resize([self.target_imgsize, self.target_imgsize], Image.BILINEAR)
image = np.array(image.convert('RGB')).astype(np.float32)
image = self.normalize_np_image(image)
return image
#reads data, preprocesses and saves to another folder with the given path.
def preprocess_and_save_images(self, dir_name, save_path=''):
preproc_folder_path = os.path.join(save_path, dir_name)
if not os.path.exists(preproc_folder_path):
os.makedirs(preproc_folder_path)
imgs_path = os.path.join(self.data_path, 'img_align_celeba/*.jpg')
print('Saving and preprocessing images ...')
for num, imgname in enumerate(glob.iglob(imgs_path, recursive=True)):
cur_image = self.load_and_preprocess_image(imgname)
cur_image = Image.fromarray(np.uint8(self.denormalize_np_image(cur_image)))
cur_image.save(preproc_folder_path + '/preprocessed_image_%d.jpg' %(num))
self.data_path= preproc_folder_path
def get_nextbatch(self, batch_size):
print("nextbatch batchsize is: ", batch_size)
assert (batch_size > 0),"Give a valid batch size"
cur_idx = 0
image_namelist = self.get_imagelist(self.data_path)
while cur_idx + batch_size <= self.num_imgs:
cur_namelist = image_namelist[cur_idx:cur_idx + batch_size]
cur_batch = [self.get_input(image_path) for image_path in cur_namelist]
cur_batch = np.array(cur_batch).astype(np.float32)
cur_idx += batch_size
yield cur_batch
def show_image(self, image, normalized=True):
if not type(image).__module__ == np.__name__:
image = image.numpy()
if normalized:
npimg = (image * 0.5) + 0.5
npimg.astype(np.uint8)
plt.imshow(npimg, interpolation='nearest')
#contains functions that load, preprocess and visualize images.
class Dataset(object):
def __init__(self, data_path, num_imgs, target_imgsize):
self.data_path = data_path
self.num_imgs = num_imgs
self.target_imgsize = target_imgsize
def normalize_np_image(self, image):
return (image / 255.0 - 0.5) / 0.5
def denormalize_np_image(self, image):
return (image * 0.5 + 0.5) * 255
def get_input(self, image_path):
image = np.array(Image.open(image_path)).astype(np.float32)
return self.normalize_np_image(image)
def get_imagelist(self, data_path, celebA=False):
if celebA == True:
imgs_path = os.path.join(data_path, 'img_align_celeba/*.jpg')
else:
imgs_path = os.path.join(data_path, '*.jpg')
all_namelist = glob.glob(imgs_path, recursive=True)
return all_namelist[:self.num_imgs]
def load_and_preprocess_image(self, image_path):
image = Image.open(image_path)
j = (image.size[0] - 100) // 2
i = (image.size[1] - 100) // 2
image = image.crop([j, i, j + 100, i + 100])
image = image.resize([self.target_imgsize, self.target_imgsize], Image.BILINEAR)
image = np.array(image.convert('RGB')).astype(np.float32)
image = self.normalize_np_image(image)
return image
#reads data, preprocesses and saves to another folder with the given path.
def preprocess_and_save_images(self, dir_name, save_path=''):
preproc_folder_path = os.path.join(save_path, dir_name)
if not os.path.exists(preproc_folder_path):
os.makedirs(preproc_folder_path)
imgs_path = os.path.join(self.data_path, 'img_align_celeba/*.jpg')
print('Saving and preprocessing images ...')
for num, imgname in enumerate(glob.iglob(imgs_path, recursive=True)):
cur_image = self.load_and_preprocess_image(imgname)
cur_image = Image.fromarray(np.uint8(self.denormalize_np_image(cur_image)))
cur_image.save(preproc_folder_path + '/preprocessed_image_%d.jpg' %(num))
self.data_path= preproc_folder_path
def get_nextbatch(self, batch_size):
assert (batch_size > 0),"Give a valid batch size"
cur_idx = 0
image_namelist = self.get_imagelist(self.data_path)
while cur_idx + batch_size <= self.num_imgs:
cur_namelist = image_namelist[cur_idx:cur_idx + batch_size]
cur_batch = [self.get_input(image_path) for image_path in cur_namelist]
cur_batch = np.array(cur_batch).astype(np.float32)
cur_idx += batch_size
yield cur_batch
def show_image(self, image, normalized=True):
if not type(image).__module__ == np.__name__:
image = image.numpy()
if normalized:
npimg = (image * 0.5) + 0.5
npimg.astype(np.uint8)
plt.imshow(npimg, interpolation='nearest')
def generator(x, args, reuse=False):
with tf.device('/gpu:0'):
with tf.variable_scope("generator", reuse=reuse):
#Layer Block 1
with tf.variable_scope("layer1"):
deconv1 = tf.layers.conv2d_transpose(inputs=x,
filters= args.n_g_filters*8,
kernel_size=4,
strides=1,
padding='valid',
use_bias=False,
name='deconv')
batch_norm1=tf.layers.batch_normalization(deconv1,
name = 'batch_norm')
relu1 = tf.nn.relu(batch_norm1, name='relu')
#Layer Block 2
with tf.variable_scope("layer2"):
deconv2 = tf.layers.conv2d_transpose(inputs=relu1,
filters=args.n_g_filters*4,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='deconv')
batch_norm2 = tf.layers.batch_normalization(deconv2,
name = 'batch_norm')
relu2 = tf.nn.relu(batch_norm2, name='relu')
#Layer Block 3
with tf.variable_scope("layer3"):
deconv3 = tf.layers.conv2d_transpose(inputs=relu2,
filters=args.n_g_filters*2,
kernel_size=4,
strides=2,
padding='same',
use_bias = False,
name='deconv')
batch_norm3 = tf.layers.batch_normalization(deconv3,
name = 'batch_norm')
relu3 = tf.nn.relu(batch_norm3, name='relu')
#Layer Block 4
with tf.variable_scope("layer4"):
deconv4 = tf.layers.conv2d_transpose(inputs=relu3,
filters=args.n_g_filters,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='deconv')
batch_norm4 = tf.layers.batch_normalization(deconv4,
name = 'batch_norm')
relu4 = tf.nn.relu(batch_norm4, name='relu')
#Output Layer
with tf.variable_scope("last_layer"):
logit = tf.layers.conv2d_transpose(inputs=relu4,
filters=3,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='logit')
output = tf.nn.tanh(logit)
return output, logit
def discriminator(x, args, reuse=False):
with tf.device('/gpu:0'):
with tf.variable_scope("discriminator", reuse=reuse):
with tf.variable_scope("layer1"):
conv1 = tf.layers.conv2d(inputs=x,
filters=args.n_f_filters,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='conv')
relu1 = tf.nn.leaky_relu(conv1, alpha=0.2, name='relu')
with tf.variable_scope("layer2"):
conv2 = tf.layers.conv2d(inputs=relu1,
filters=args.n_f_filters*2,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='conv')
batch_norm2 = tf.layers.batch_normalization(conv2,name='batch_norm')
relu2 = tf.nn.leaky_relu(batch_norm2, alpha=0.2, name='relu')
with tf.variable_scope("layer3"):
conv3 = tf.layers.conv2d(inputs=relu2,
filters=args.n_f_filters*4,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='conv')
batch_norm3 = tf.layers.batch_normalization(conv3, name='batch_norm')
relu3 = tf.nn.leaky_relu(batch_norm3, name='relu')
with tf.variable_scope("layer4"):
conv4 = tf.layers.conv2d(inputs=relu3,
filters=args.n_f_filters*8,
kernel_size=4,
strides=2,
padding='same',
use_bias=False,
name='conv')
batch_norm4 = tf.layers.batch_normalization(conv4, name='batch_norm')
relu4 = tf.nn.leaky_relu(batch_norm4, alpha=0.2, name='relu')
with tf.variable_scope("last_layer"):
logit = tf.layers.conv2d(inputs=relu4,
filters=1,
kernel_size=4,
strides=1,
padding='valid',
use_bias=False,
name='conv')
output = tf.nn.sigmoid(logit)
return output, logit
def sample_z(dim_z, num_batch):
mu = 0
sigma = 1
s = np.random.normal(mu, sigma, num_batch*dim_z)
samples = s.reshape(num_batch, 1, 1, dim_z)
##dist = tf.distributions.Normal(0.0, 1.0)
##samples = dist.sample([num_batch, 1, 1, dim_z])
return samples
#64,1,1,100 6400
sample_z(100, 64)
def get_losses(d_real_logits, d_fake_logits):
#add new loss function here
###d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real_logits, labels=tf.ones_like(d_real_logits)))
###d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits, labels=tf.zeros_like(d_fake_logits)))
###d_loss = d_loss_real + d_loss_fake
###g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits, labels=tf.ones_like(d_fake_logits)))
###return d_loss, g_loss
d_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real_logits,labels=tf.ones_like(d_real_logits)) + tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits,labels=tf.zeros_like(d_fake_logits)))
g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logits,labels=tf.ones_like(d_fake_logits)))
return d_loss, g_loss
def get_optimizers(learning_rate, beta1, beta2):
d_optimizer = tf.train.AdamOptimizer(learning_rate, beta1, beta2)
g_optimizer = tf.train.AdamOptimizer(learning_rate, beta1, beta2)
return d_optimizer, g_optimizer
def optimize(d_optimizer, g_optimizer, d_loss, g_loss):
d_step = d_optimizer.minimize(d_loss)
g_step = g_optimizer.minimize(g_loss)
return d_step, g_step
LOGDIR = "logs_basic_dcgan"
def merge_images(image_batch, size):
h,w = image_batch.shape[1], image_batch.shape[2]
c = image_batch.shape[3]
img = np.zeros((int(h*size[0]), w*size[1], c))
for idx, im in enumerate(image_batch):
i = idx % size[1]
j = idx // size[1]
img[j*h:j*h+h, i*w:i*w+w,:] = im
return img
itr_fh = open('basic_gan_itr.txt', 'a+')
def train(args):
tf.reset_default_graph()
data_loader = Dataset(args.data_path, args.num_images, args.image_size)
#data_loader.preprocess_and_save_images('preprocessed', 'results_celebA') #preprocess the images once
X = tf.placeholder(tf.float32, shape=[args.batch_size, args.image_size , args.image_size, 3])
Z = tf.placeholder(tf.float32, shape=[args.batch_size, 1, 1, args.dim_z])
G_sample, _ = generator(Z, args)
D_real, D_real_logits = discriminator(X, args)
D_fake, D_fake_logits = discriminator(G_sample, args, reuse=True)
d_loss, g_loss = get_losses(D_real_logits, D_fake_logits)
d_optimizer, g_optimizer = get_optimizers(args.lr, args.beta1, args.beta2)
d_step, g_step = optimize(d_optimizer, g_optimizer, d_loss, g_loss)
###z_sum = tf.summary.histogram('z', Z)
###d_sum = tf.summary.histogram('d', D_real)
###G_sum = tf.summary.histogram('g', G_sample)
###d_loss_sum = tf.summary.scalar('d_loss', d_loss)
###g_loss_sum = tf.summary.scalar('g_loss', g_loss)
###d_sum = tf.summary.merge([z_sum, d_sum, d_loss_sum])
###g_sum = tf.summary.merge([z_sum, G_sum, g_loss_sum])
###saver = tf.train.Saver()
###merged_summary = tf.summary.merge_all()
###d_loss_summary = tf.summary.scalar("Discriminator_Total_Loss", d_loss)
###g_loss_summary = tf.summary.scalar("Generator_Total_Loss", g_loss)
###merged_summary = tf.summary.merge_all()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(args.n_epoch):
for itr, real_batch in enumerate(data_loader.get_nextbatch(args.batch_size)):
print('itr is %d, and epoch is %d' %(itr, epoch))
itr_fh.write("epoch: " + str(epoch) + " itr: " + str(itr) + "\n")
Z_sample = sample_z(args.dim_z, args.batch_size)
_, _ = sess.run([d_step, g_step], feed_dict={X:real_batch , Z:Z_sample})
sample = sess.run(G_sample, feed_dict={Z:Z_sample})
print("sample size is: ", sample.shape)
if itr==3164: #num_images/batch_size
im_merged = merge_images(sample[:16], [4,4])
plt.imsave('sample_gan_images/im_merged_epoch_%d.png' %(epoch), im_merged )
scipy.misc.imsave('sample_gan_images/im_epoch_%d_itr_%d.png' %(epoch,itr), sample[1])
##merged_summary = sess.run(merged_summary, feed_dict={X:real_batch , Z:Z_sample})
###writer = tf.summary.FileWriter(LOGDIR)
###writer.add_summary(merged_summary, itr)
###d_loss_summary = tf.summary.scalar("Discriminator_Total_Loss", d_loss)
###g_loss_summary = tf.summary.scalar("Generator_Total_Loss", g_loss)
###merged_summary = tf.summary.merge_all()
###writer.add_graph(sess.graph)
###saver.save(sess, save_path='logs_basic_dcgan/gan.ckpt')
train(args)
这是在前 5 个 epoch 结束时创建的图像。我也评论了与 tensorboard 相关的东西，因为不幸的是它使它变得非常慢。
纪元 0 结束：
时代 1 结束：
时代 2 结束：
纪元 3 结束：
时代 4 结束：
我认为问题与优化器的定义有关：
def optimize(d_optimizer, g_optimizer, d_loss, g_loss):
d_step = d_optimizer.minimize(d_loss)
g_step = g_optimizer.minimize(g_loss)
return d_step, g_step
虽然您用相应的损失定义了每个优化器，但您没有传递每个优化器将训练的变量列表。因此，通过 default 函数 minimize 将考虑图形集合 GraphKeys.TRAINABLE_VARIABLES 下的所有变量。由于您的所有变量都在此图形集合下定义，因此当您调用 d_step 和调用 g_step.
您必须为每个模型定义变量列表。由于您使用的是变量范围，因此一种方法是：
def optimize(d_optimizer, g_optimizer, d_loss, g_loss):
t_vars = tf.trainable_variables()
d_vars = [var for var in t_vars if var.name.startswith('discriminator')]
g_vars = [var for var in t_vars if var.name.startswith('generator')]
d_step = d_optimizer.minimize(d_loss, var_list=d_vars)
g_step = g_optimizer.minimize(g_loss, var_list=g_vars)
return d_step, g_step