pythontensorflowdeep-learninggenerative-adversarial-network
DCGAN Tensorflow code doesn't produce faces on CelebA dataset
I have written the following code but does not produce faces on the celebA dataset. I think it should create some sort of face (even if very blurry) at the last iteration of each epoch. However, it just creates noisy squares with no visible face. I am quite new to GANs and I am not sure how to debug this Deep Convolutional GAN (DCGAN) in order to figure what's going wrong.
My code might be easier to be seen here: https://pastebin.com/c4QUqxJy Here is the code:
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)
Here is the images created at the end of first 5 epochs. I also have commented stuff related to tensorboard because it makes it very slow unfortunately.
end of Epoch 0:
end of Epoch 1:
end of Epoch 2:
end of Epoch 3:
end of Epoch 4:
Solution
I think the problem is related with the definition of the optimizers:
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
Although you define each optimizer with the corresponding loss, you are not passing the list of variables that will be trained by each optimizer. Therefore, by default the function minimize will consider all variables under the graph collection GraphKeys.TRAINABLE_VARIABLES. Since all your variables are defined under this graph collection, your current code actually updates all variables from the generator and from the discriminator when you call d_step and when you call g_step.
You have to define the list of variables for each model. Since you are using variable scopes, one way to do that is:
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