MNIST GAN generators white area in middle surrounded by black - tensorflow

Question

The following code is copied from a GAN MNIST tutorial on UDEMY. When I run the code, it converges towards creating images with a large white area in the center that is black at the sides (picture an empty filled circle against a black background). I have no idea what the problem is as I have only done what the tutorial told me to do word for word. The only difference is that I extract the MNIST data differently. Is there something about tensorflow that has changed recently?
import tensorflow as tf
import numpy as np
import gzip
from PIL import Image
import os.path
def extract_data(filename, num_images):
"""Extract the images into a 4D tensor [image index, y, x, channels].
Values are rescaled from [0, 255] down to [-0.5, 0.5].
"""
print('Extracting', filename)
with gzip.open(filename) as bytestream:
bytestream.read(16)
buf = bytestream.read(28 * 28 * num_images)
data = np.frombuffer(buf, dtype=np.uint8).astype(np.float32)
#data = (data - (PIXEL_DEPTH / 2.0)) / PIXEL_DEPTH
data = data.reshape(num_images, 28, 28, 1)
return data
fname_img_train = extract_data('../Data/MNIST/train-images-idx3-ubyte.gz', 60000)
def generator(z, reuse=None):
with tf.variable_scope('gen',reuse=reuse):
hidden1 = tf.layers.dense(inputs=z,units=128)
alpha = 0.01
hidden1=tf.maximum(alpha*hidden1,hidden1)
hidden2=tf.layers.dense(inputs=hidden1,units=128)
hidden2 = tf.maximum(alpha*hidden2,hidden2)
output=tf.layers.dense(hidden2,units=784, activation=tf.nn.tanh)
return output
def discriminator(X, reuse=None):
with tf.variable_scope('dis',reuse=reuse):
hidden1=tf.layers.dense(inputs=X,units=128)
alpha=0.01
hidden1=tf.maximum(alpha*hidden1,hidden1)
hidden2=tf.layers.dense(inputs=hidden1,units=128)
hidden2=tf.maximum(alpha*hidden2,hidden2)
logits=tf.layers.dense(hidden2,units=1)
output=tf.sigmoid(logits)
return output, logits
real_images=tf.placeholder(tf.float32,shape=[None,784])
z=tf.placeholder(tf.float32,shape=[None,100])
G = generator(z)
D_output_real, D_logits_real = discriminator(real_images)
D_output_fake, D_logits_fake = discriminator(G,reuse=True)
def loss_func(logits_in,labels_in):
return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
logits=logits_in,labels=labels_in))
D_real_loss = loss_func(D_logits_real,tf.ones_like(D_logits_real)*0.9)
D_fake_loss = loss_func(D_logits_fake,tf.zeros_like(D_logits_real))
D_loss = D_real_loss + D_fake_loss
G_loss = loss_func(D_logits_fake,tf.ones_like(D_logits_fake))
learning_rate = 0.001
tvars = tf.trainable_variables()
d_vars= [var for var in tvars if 'dis' in var.name]
g_vars = [var for var in tvars if 'gen' in var.name]
D_trainer = tf.train.AdamOptimizer(learning_rate).minimize(D_loss,var_list=d_vars)
G_trainer = tf.train.AdamOptimizer(learning_rate).minimize(G_loss,var_list=g_vars)
batch_size=100
epochs=30
set_size=60000
init = tf.global_variables_initializer()
samples=[]
def create_image(img, name):
img = np.reshape(img, (28, 28))
print("before")
print(img)
img = (np.multiply(np.divide(np.add(img, 1.0), 2.0),255.0).astype(np.int16))
print("after")
print(img)
im = Image.fromarray(img.astype('uint8'))
im.save(name)
with tf.Session() as sess:
sess.run(init)
for epoch in range(epochs):
np.random.shuffle(fname_img_train)
num_batches=int(set_size/batch_size)
for i in range(num_batches):
batch = fname_img_train[i*batch_size:((i+1)*batch_size)]
batch_images = np.reshape(batch, (batch_size,784))
batch_images = batch_images*2.0-1.0
batch_z = np.random.uniform(-1,1,size=(batch_size,100))
_ = sess.run(D_trainer, feed_dict={real_images:batch_images,z:batch_z})
_ = sess.run(G_trainer,feed_dict={z:batch_z})
print("ON EPOCH {}".format(epoch))
sample_z = np.random.uniform(-1,1,size=(batch_size,100))
gen_sample = sess.run(G,feed_dict={z:sample_z})
create_image(gen_sample[0], "img"+str(epoch)+".png")

Answer 1

As far as I can see, you are not normalizing the training data. Instead of using your extract_data() function, it is much easier to do the following:
from tensorflow.keras.datasets.mnist import load_data()
(train_data, train_labels), _ = load_data()
train_data /= 255.
Besides, usually people sample twice from the latent space each epoch: once for the discriminator and once for the generator. Still, it did not seem to make a difference.
After implementing these changes, using a batch size of 200 and training for 100 epochs, I got the following result: gen_sample. The result is pretty bad, but it is definitely better than an "empty filled circle against a black background".
Note that the architecture of the generator and of the discriminator that you are using is very simple. From my experience, stacking some convolutional layers gives perfect results. In addition, I would not use the tf.maximum() function, since it creates discontinuities that may negatively impact the flow of the gradients.
Finally, instead of your create_image() function, I used the following:
def plot_mnist(samples, name):
fig = plt.figure(figsize=(6,6))
gs = gridspec.GridSpec(6,6)
gs.update(wspace=0.05, hspace=0.05)
for i, sample in enumerate(samples):
ax = plt.subplot(gs[i])
plt.axis('off')
ax.set_xticklabels([])
ax.set_yticklabels([])
ax.set_aspect('equal')
plt.imshow(sample.reshape(28,28), cmap='Greys_r')
plt.savefig('{}.png'.format(name))
plt.close()
There are many different ways of improving the quality of a GAN model, and the majority of those techniques can be easily found online. Please let me know if you have any specific question.