Generative Adversarial Network

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 [1]:
%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')


Extracting MNIST_data\train-images-idx3-ubyte.gz
Extracting MNIST_data\train-labels-idx1-ubyte.gz
Extracting MNIST_data\t10k-images-idx3-ubyte.gz
Extracting MNIST_data\t10k-labels-idx1-ubyte.gz

Model Inputs

First we need to create the inputs for our graph. We need two inputs, one for the discriminator and one for the generator. Here we'll call the discriminator input inputs_real and the generator input inputs_z. We'll assign them the appropriate sizes for each of the networks.


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

Generator network

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.

Variable Scope

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.

Leaky ReLU

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: $$ f(x) = max(\alpha * x, x) $$

Tanh Output

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.


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, activation=None)
        out = tf.tanh(logits)
        
        return out

Discriminator

The discriminator network is almost exactly the same as the generator network, except that we're using a sigmoid output layer.


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

Hyperparameters


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

Build network

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 [7]:
tf.reset_default_graph()
# Create our input placeholders
input_real, input_z = model_inputs(input_size, z_size)

# Build the model
g_model = generator(input_z, input_size, n_units=g_hidden_size, alpha=alpha)
# g_model is the generator output

d_model_real, d_logits_real = discriminator(input_real, n_units=d_hidden_size, alpha=alpha)
d_model_fake, d_logits_fake = discriminator(g_model, reuse=True, n_units=d_hidden_size, alpha=alpha)

Discriminator and Generator Losses

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 [8]:
# 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)))

Optimizers

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 [9]:
# 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)

Training


In [10]:
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, n_units=g_hidden_size, reuse=True, alpha=alpha),
                       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)


Epoch 1/100... Discriminator Loss: 0.3717... Generator Loss: 3.8729
Epoch 2/100... Discriminator Loss: 0.5164... Generator Loss: 2.3166
Epoch 3/100... Discriminator Loss: 0.4797... Generator Loss: 3.7419
Epoch 4/100... Discriminator Loss: 0.8515... Generator Loss: 2.5611
Epoch 5/100... Discriminator Loss: 1.2399... Generator Loss: 4.6831
Epoch 6/100... Discriminator Loss: 1.0274... Generator Loss: 3.3913
Epoch 7/100... Discriminator Loss: 0.7801... Generator Loss: 2.7318
Epoch 8/100... Discriminator Loss: 0.6282... Generator Loss: 2.6684
Epoch 9/100... Discriminator Loss: 1.4837... Generator Loss: 2.3605
Epoch 10/100... Discriminator Loss: 2.1502... Generator Loss: 1.4087
Epoch 11/100... Discriminator Loss: 1.0172... Generator Loss: 2.6197
Epoch 12/100... Discriminator Loss: 1.4271... Generator Loss: 2.2649
Epoch 13/100... Discriminator Loss: 1.0086... Generator Loss: 1.7981
Epoch 14/100... Discriminator Loss: 1.3969... Generator Loss: 1.7358
Epoch 15/100... Discriminator Loss: 1.7089... Generator Loss: 2.2373
Epoch 16/100... Discriminator Loss: 1.6251... Generator Loss: 1.5852
Epoch 17/100... Discriminator Loss: 1.1822... Generator Loss: 2.1829
Epoch 18/100... Discriminator Loss: 1.3160... Generator Loss: 2.0353
Epoch 19/100... Discriminator Loss: 0.8703... Generator Loss: 2.0834
Epoch 20/100... Discriminator Loss: 1.1460... Generator Loss: 2.1618
Epoch 21/100... Discriminator Loss: 1.2032... Generator Loss: 1.4160
Epoch 22/100... Discriminator Loss: 1.3753... Generator Loss: 1.9852
Epoch 23/100... Discriminator Loss: 1.3630... Generator Loss: 1.3455
Epoch 24/100... Discriminator Loss: 0.9100... Generator Loss: 1.7652
Epoch 25/100... Discriminator Loss: 1.3816... Generator Loss: 1.3716
Epoch 26/100... Discriminator Loss: 0.9868... Generator Loss: 1.7875
Epoch 27/100... Discriminator Loss: 1.0643... Generator Loss: 1.8689
Epoch 28/100... Discriminator Loss: 1.4859... Generator Loss: 1.8740
Epoch 29/100... Discriminator Loss: 1.2271... Generator Loss: 1.6898
Epoch 30/100... Discriminator Loss: 1.0608... Generator Loss: 2.1647
Epoch 31/100... Discriminator Loss: 1.2227... Generator Loss: 1.9518
Epoch 32/100... Discriminator Loss: 1.0394... Generator Loss: 1.9031
Epoch 33/100... Discriminator Loss: 0.8694... Generator Loss: 2.8467
Epoch 34/100... Discriminator Loss: 0.9722... Generator Loss: 1.9111
Epoch 35/100... Discriminator Loss: 0.8600... Generator Loss: 2.3834
Epoch 36/100... Discriminator Loss: 1.1989... Generator Loss: 1.9588
Epoch 37/100... Discriminator Loss: 0.9685... Generator Loss: 1.9430
Epoch 38/100... Discriminator Loss: 1.0744... Generator Loss: 1.6864
Epoch 39/100... Discriminator Loss: 1.0853... Generator Loss: 2.2258
Epoch 40/100... Discriminator Loss: 0.9280... Generator Loss: 1.8204
Epoch 41/100... Discriminator Loss: 0.9255... Generator Loss: 1.9563
Epoch 42/100... Discriminator Loss: 1.1604... Generator Loss: 2.6429
Epoch 43/100... Discriminator Loss: 0.9511... Generator Loss: 1.8912
Epoch 44/100... Discriminator Loss: 1.0524... Generator Loss: 1.4524
Epoch 45/100... Discriminator Loss: 1.1625... Generator Loss: 2.1108
Epoch 46/100... Discriminator Loss: 0.8632... Generator Loss: 2.2580
Epoch 47/100... Discriminator Loss: 1.2043... Generator Loss: 1.8506
Epoch 48/100... Discriminator Loss: 0.9833... Generator Loss: 1.8992
Epoch 49/100... Discriminator Loss: 1.0178... Generator Loss: 2.4039
Epoch 50/100... Discriminator Loss: 0.8084... Generator Loss: 2.1072
Epoch 51/100... Discriminator Loss: 0.9347... Generator Loss: 1.8241
Epoch 52/100... Discriminator Loss: 0.8394... Generator Loss: 2.1113
Epoch 53/100... Discriminator Loss: 1.0428... Generator Loss: 1.9732
Epoch 54/100... Discriminator Loss: 0.9429... Generator Loss: 1.6812
Epoch 55/100... Discriminator Loss: 1.0712... Generator Loss: 1.3454
Epoch 56/100... Discriminator Loss: 1.1381... Generator Loss: 1.9971
Epoch 57/100... Discriminator Loss: 0.9655... Generator Loss: 1.7299
Epoch 58/100... Discriminator Loss: 1.2137... Generator Loss: 1.5399
Epoch 59/100... Discriminator Loss: 1.0022... Generator Loss: 1.4885
Epoch 60/100... Discriminator Loss: 0.9987... Generator Loss: 1.8311
Epoch 61/100... Discriminator Loss: 0.9332... Generator Loss: 1.8154
Epoch 62/100... Discriminator Loss: 1.0647... Generator Loss: 1.8425
Epoch 63/100... Discriminator Loss: 0.9881... Generator Loss: 1.6489
Epoch 64/100... Discriminator Loss: 1.0009... Generator Loss: 1.6786
Epoch 65/100... Discriminator Loss: 0.8809... Generator Loss: 1.8859
Epoch 66/100... Discriminator Loss: 0.9466... Generator Loss: 1.7459
Epoch 67/100... Discriminator Loss: 1.0993... Generator Loss: 1.5636
Epoch 68/100... Discriminator Loss: 0.9865... Generator Loss: 1.7583
Epoch 69/100... Discriminator Loss: 1.0176... Generator Loss: 1.8221
Epoch 70/100... Discriminator Loss: 1.0723... Generator Loss: 2.0352
Epoch 71/100... Discriminator Loss: 1.1983... Generator Loss: 1.6830
Epoch 72/100... Discriminator Loss: 1.0846... Generator Loss: 2.3876
Epoch 73/100... Discriminator Loss: 0.9848... Generator Loss: 1.8892
Epoch 74/100... Discriminator Loss: 1.1447... Generator Loss: 1.6627
Epoch 75/100... Discriminator Loss: 1.0517... Generator Loss: 1.9052
Epoch 76/100... Discriminator Loss: 0.8079... Generator Loss: 2.7591
Epoch 77/100... Discriminator Loss: 0.8899... Generator Loss: 1.8953
Epoch 78/100... Discriminator Loss: 1.1558... Generator Loss: 1.6257
Epoch 79/100... Discriminator Loss: 0.9353... Generator Loss: 1.9043
Epoch 80/100... Discriminator Loss: 1.2365... Generator Loss: 1.5800
Epoch 81/100... Discriminator Loss: 0.8318... Generator Loss: 2.1138
Epoch 82/100... Discriminator Loss: 1.1416... Generator Loss: 2.0404
Epoch 83/100... Discriminator Loss: 1.0485... Generator Loss: 1.7069
Epoch 84/100... Discriminator Loss: 1.0978... Generator Loss: 1.2435
Epoch 85/100... Discriminator Loss: 1.0328... Generator Loss: 1.6038
Epoch 86/100... Discriminator Loss: 1.1044... Generator Loss: 2.2230
Epoch 87/100... Discriminator Loss: 1.0547... Generator Loss: 1.4468
Epoch 88/100... Discriminator Loss: 0.9236... Generator Loss: 1.9706
Epoch 89/100... Discriminator Loss: 1.0055... Generator Loss: 2.1177
Epoch 90/100... Discriminator Loss: 1.1323... Generator Loss: 1.6080
Epoch 91/100... Discriminator Loss: 0.8968... Generator Loss: 1.7801
Epoch 92/100... Discriminator Loss: 1.0255... Generator Loss: 1.5662
Epoch 93/100... Discriminator Loss: 1.1016... Generator Loss: 1.8286
Epoch 94/100... Discriminator Loss: 1.1601... Generator Loss: 1.3521
Epoch 95/100... Discriminator Loss: 1.0180... Generator Loss: 1.8499
Epoch 96/100... Discriminator Loss: 1.2264... Generator Loss: 2.0485
Epoch 97/100... Discriminator Loss: 0.9995... Generator Loss: 1.8882
Epoch 98/100... Discriminator Loss: 1.0634... Generator Loss: 1.4197
Epoch 99/100... Discriminator Loss: 1.1576... Generator Loss: 1.3579
Epoch 100/100... Discriminator Loss: 1.0401... Generator Loss: 1.7269

Training loss

Here we'll check out the training losses for the generator and discriminator.


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]:
<matplotlib.legend.Legend at 0x7f4e8947fbe0>

Generator samples from training

Here we can view samples of images from the generator. First we'll look at images taken while training.


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.

Sampling from the generator

We can also get completely new images from the generator by using the checkpoint we saved after training. We just need to pass in a new latent vector $z$ and we'll get new samples!


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, n_units=g_hidden_size, reuse=True, alpha=alpha),
                   feed_dict={input_z: sample_z})
_ = view_samples(0, [gen_samples])