In this notebook, we'll be building a generative adversarial network (GAN) trained on the MNIST dataset. From this, we'll be able to generate new handwritten digits!
GANs were first reported on in 2014 from Ian Goodfellow and others in Yoshua Bengio's lab. Since then, GANs have exploded in popularity. Here are a few examples to check out:
The idea behind GANs is that you have two networks, a generator $G$ and a discriminator $D$, competing against each other. The generator makes fake data to pass to the discriminator. The discriminator also sees real data and predicts if the data it's received is real or fake. The generator is trained to fool the discriminator, it wants to output data that looks as close as possible to real data. And the discriminator is trained to figure out which data is real and which is fake. What ends up happening is that the generator learns to make data that is indistiguishable from real data to the discriminator.
The general structure of a GAN is shown in the diagram above, using MNIST images as data. The latent sample is a random vector the generator uses to contruct it's fake images. As the generator learns through training, it figures out how to map these random vectors to recognizable images that can foold the discriminator.
The output of the discriminator is a sigmoid function, where 0 indicates a fake image and 1 indicates an real image. If you're interested only in generating new images, you can throw out the discriminator after training. Now, let's see how we build this thing in TensorFlow.
In [3]:
%matplotlib inline
import pickle as pkl
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
In [2]:
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data')
In [3]:
def model_inputs(real_dim, z_dim):
inputs_real = tf.placeholder(tf.float32, (None, real_dim), name='input_real')
inputs_z = tf.placeholder(tf.float32, (None, z_dim), name='input_z')
return inputs_real, inputs_z
Here we'll build the generator network. To make this network a universal function approximator, we'll need at least one hidden layer. We should use a leaky ReLU to allow gradients to flow backwards through the layer unimpeded. A leaky ReLU is like a normal ReLU, except that there is a small non-zero output for negative input values.
Here we need to use tf.variable_scope for two reasons. Firstly, we're going to make sure all the variable names start with generator. Similarly, we'll prepend discriminator to the discriminator variables. This will help out later when we're training the separate networks.
We could just use tf.name_scope to set the names, but we also want to reuse these networks with different inputs. For the generator, we're going to train it, but also sample from it as we're training and after training. The discriminator will need to share variables between the fake and real input images. So, we can use the reuse keyword for tf.variable_scope to tell TensorFlow to reuse the variables instead of creating new ones if we build the graph again.
To use tf.variable_scope, you use a with statement:
with tf.variable_scope('scope_name', reuse=False):
# code here
Here's more from the TensorFlow documentation to get another look at using tf.variable_scope.
TensorFlow doesn't provide an operation for leaky ReLUs, so we'll need to make one . For this you can use take the outputs from a linear fully connected layer and pass them to tf.maximum. Typically, a parameter alpha sets the magnitude of the output for negative values. So, the output for negative input (x) values is alpha*x, and the output for positive x is x:
$$
f(x) = max(\alpha * x, x)
$$
The generator has been found to perform the best with $tanh$ for the generator output. This means that we'll have to rescale the MNIST images to be between -1 and 1, instead of 0 and 1. Along with the $tanh$ output, we also need to return the logits for use in calculating the loss with tf.nn.sigmoid_cross_entropy_with_logits.
In [4]:
def generator(z, out_dim, n_units=128, reuse=False, alpha=0.01):
with tf.variable_scope('generator', reuse=reuse):
# Hidden layer
h1 = tf.layers.dense(z, n_units, activation=None)
# Leaky ReLU
h1 = tf.maximum(alpha * h1, h1)
# Logits and tanh output
logits = tf.layers.dense(h1, out_dim)
out = tf.tanh(logits)
return out, logits
In [5]:
def discriminator(x, n_units=128, reuse=False, alpha=0.01):
with tf.variable_scope('discriminator', reuse=reuse):
# Hidden layer
h1 = tf.layers.dense(x, n_units, activation=None)
# Leaky ReLU
h1 = tf.maximum(alpha * h1, h1)
logits = tf.layers.dense(h1, 1, activation=None)
out = tf.sigmoid(logits)
return out, logits
In [6]:
# Size of input image to discriminator
input_size = 784
# Size of latent vector to generator
z_size = 100
# Sizes of hidden layers in generator and discriminator
g_hidden_size = 128
d_hidden_size = 128
# Leak factor for leaky ReLU
alpha = 0.01
# Smoothing
smooth = 0.1
Now we're building the network from the functions defined above.
First is to get our inputs, input_real, input_z from model_inputs using the sizes of the input and z.
Then, we'll create the generator, generator(input_z, input_size). This builds the generator with the appropriate input and output sizes.
Then the discriminators. We'll build two of them, one for real data and one for fake data. Since we want the weights to be the same for both real and fake data, we need to reuse the variables. For the fake data, we're getting it from the generator as g_model. So the real data discriminator is discriminator(input_real) while the fake discriminator is discriminator(g_model, reuse=True).
In [13]:
tf.reset_default_graph()
# Create our input placeholders
input_real, input_z = model_inputs(input_size, z_size)
# Build the model
g_model, g_logits = generator(input_z, input_size)
# g_model is the generator output
d_model_real, d_logits_real = discriminator(input_real)
d_model_fake, d_logits_fake = discriminator(g_model, reuse=True)
Now we need to calculate the losses, which is a little tricky. For the discriminator, the total loss is the sum of the losses for real and fake images, d_loss = d_loss_real + d_loss_fake. The losses will by sigmoid cross-entropys, which we can get with tf.nn.sigmoid_cross_entropy_with_logits. We'll also wrap that in tf.reduce_mean to get the mean for all the images in the batch. So the losses will look something like
tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=labels))
For the real image logits, we'll use d_logits_real which we got from the discriminator in the cell above. For the labels, we want them to be all ones, since these are all real images. To help the discriminator generalize better, the labels are reduced a bit from 1.0 to 0.9, for example, using the parameter smooth. This is known as label smoothing, typically used with classifiers to improve performance. In TensorFlow, it looks something like labels = tf.ones_like(tensor) * (1 - smooth)
The discriminator loss for the fake data is similar. The logits are d_logits_fake, which we got from passing the generator output to the discriminator. These fake logits are used with labels of all zeros. Remember that we want the discriminator to output 1 for real images and 0 for fake images, so we need to set up the losses to reflect that.
Finally, the generator losses are using d_logits_fake, the fake image logits. But, now the labels are all ones. The generator is trying to fool the discriminator, so it wants to discriminator to output ones for fake images.
In [14]:
# Calculate losses
d_loss_real = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_real,
labels=tf.ones_like(d_logits_real) * (1 - smooth)))
d_loss_fake = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake,
labels=tf.zeros_like(d_logits_real)))
d_loss = d_loss_real + d_loss_fake
g_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake,
labels=tf.ones_like(d_logits_fake)))
We want to update the generator and discriminator variables separately. So we need to get the variables for each part build optimizers for the two parts. To get all the trainable variables, we use tf.trainable_variables(). This creates a list of all the variables we've defined in our graph.
For the generator optimizer, we only want to generator variables. Our past selves were nice and used a variable scope to start all of our generator variable names with generator. So, we just need to iterate through the list from tf.trainable_variables() and keep variables to start with generator. Each variable object has an attribute name which holds the name of the variable as a string (var.name == 'weights_0' for instance).
We can do something similar with the discriminator. All the variables in the discriminator start with discriminator.
Then, in the optimizer we pass the variable lists to var_list in the minimize method. This tells the optimizer to only update the listed variables. Something like tf.train.AdamOptimizer().minimize(loss, var_list=var_list) will only train the variables in var_list.
In [15]:
# Optimizers
learning_rate = 0.002
# Get the trainable_variables, split into G and D parts
t_vars = tf.trainable_variables()
g_vars = [var for var in t_vars if var.name.startswith('generator')]
d_vars = [var for var in t_vars if var.name.startswith('discriminator')]
d_train_opt = tf.train.AdamOptimizer(learning_rate).minimize(d_loss, var_list=d_vars)
g_train_opt = tf.train.AdamOptimizer(learning_rate).minimize(g_loss, var_list=g_vars)
In [16]:
!mkdir checkpoints
In [ ]:
batch_size = 100
epochs = 100
samples = []
losses = []
# Only save generator variables
saver = tf.train.Saver(var_list=g_vars)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for e in range(epochs):
for ii in range(mnist.train.num_examples//batch_size):
batch = mnist.train.next_batch(batch_size)
# Get images, reshape and rescale to pass to D
batch_images = batch[0].reshape((batch_size, 784))
batch_images = batch_images*2 - 1
# Sample random noise for G
batch_z = np.random.uniform(-1, 1, size=(batch_size, z_size))
# Run optimizers
_ = sess.run(d_train_opt, feed_dict={input_real: batch_images, input_z: batch_z})
_ = sess.run(g_train_opt, feed_dict={input_z: batch_z})
# At the end of each epoch, get the losses and print them out
train_loss_d = sess.run(d_loss, {input_z: batch_z, input_real: batch_images})
train_loss_g = g_loss.eval({input_z: batch_z})
print("Epoch {}/{}...".format(e+1, epochs),
"Discriminator Loss: {:.4f}...".format(train_loss_d),
"Generator Loss: {:.4f}".format(train_loss_g))
# Save losses to view after training
losses.append((train_loss_d, train_loss_g))
# Sample from generator as we're training for viewing afterwards
sample_z = np.random.uniform(-1, 1, size=(16, z_size))
gen_samples, _ = sess.run(
generator(input_z, input_size, reuse=True),
feed_dict={input_z: sample_z})
samples.append(gen_samples)
saver.save(sess, './checkpoints/generator.ckpt')
# Save training generator samples
with open('train_samples.pkl', 'wb') as f:
pkl.dump(samples, f)
In [19]:
fig, ax = plt.subplots()
losses = np.array(losses)
plt.plot(losses.T[0], label='Discriminator')
plt.plot(losses.T[1], label='Generator')
plt.title("Training Losses")
plt.legend()
Out[19]:
In [4]:
def view_samples(epoch, samples):
fig, axes = plt.subplots(figsize=(7,7), nrows=4, ncols=4, sharey=True, sharex=True)
for ax, img in zip(axes.flatten(), samples[epoch]):
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
im = ax.imshow(img.reshape((28,28)), cmap='Greys_r')
return fig, axes
In [5]:
# Load samples from generator taken while training
with open('train_samples.pkl', 'rb') as f:
samples = pkl.load(f)
These are samples from the final training epoch. You can see the generator is able to reproduce numbers like 1, 7, 3, 2. Since this is just a sample, it isn't representative of the full range of images this generator can make.
In [8]:
_ = view_samples(-1, samples)
Below I'm showing the generated images as the network was training, every 10 epochs. With bonus optical illusion!
In [23]:
rows, cols = 10, 6
fig, axes = plt.subplots(figsize=(7,12), nrows=rows, ncols=cols, sharex=True, sharey=True)
for sample, ax_row in zip(samples[::int(len(samples)/rows)], axes):
for img, ax in zip(sample[::int(len(sample)/cols)], ax_row):
ax.imshow(img.reshape((28,28)), cmap='Greys_r')
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
It starts out as all noise. Then it learns to make only the center white and the rest black. You can start to see some number like structures appear out of the noise like 1s and 9s.
In [32]:
saver = tf.train.Saver(var_list=g_vars)
with tf.Session() as sess:
saver.restore(sess, tf.train.latest_checkpoint('checkpoints'))
sample_z = np.random.uniform(-1, 1, size=(16, z_size))
gen_samples, _ = sess.run(
generator(input_z, input_size, reuse=True),
feed_dict={input_z: sample_z})
_ = view_samples(0, [gen_samples])