$29
Programming Assignment 4: CycleGAN Solution
Introduction
In this assignment, you’ll get hands-on experience coding and training GANs. This assignment is divided into two parts: in the rst part, we will implement a speci c type of GAN designed to process images, called a Deep Convolutional GAN (DCGAN). We’ll train the DCGAN to generate emojis from samples of random noise. In the second part, we will implement a more complex GAN architecture called CycleGAN, which was designed for the task of image-to-image translation (described in more detail in Part 2). We’ll train the CycleGAN to convert between Apple-style and Windows-style emojis.
In both parts, you’ll gain experience implementing GANs by writing code for the generator, discriminator, and training loop, for each model.
Important Note
We provide a script to check your models, that you can run as follows:
python model_checker.py
This checks that the outputs of DCGenerator, DCDiscriminator, and CycleGenerator match the expected outputs for speci c inputs. This model checker is provided for convenience only. It may give false negatives, so do not use it as the only way to check that your model is correct.
The training scripts run much faster ( 5x faster) on the teaching lab machines if you add
MKL NUM THREADS=1 before the Python call, as follows:
MKL NUM THREADS=1 python cycle gan.py --load=pretrained cycle --train iters=100
Remember to add MKL NUM THREADS=1 before each Python command in this assignment.
https://markus.teach.cs.toronto.edu/csc321-2018-01
http://www.cs.toronto.edu/~rgrosse/courses/csc321_2018/syllabus.pdf
1
CSC321 Programming Assignment 4
Part 1: Deep Convolutional GAN (DCGAN)
For the rst part of this assignment, we will implement a Deep Convolutional GAN (DCGAN). A DCGAN is simply a GAN that uses a convolutional neural network as the discriminator, and a network composed of transposed convolutions as the generator. To implement the DCGAN, we need to specify three things: 1) the generator, 2) the discriminator, and 3) the training procedure. We will develop each of these three components in the following subsections.
Implement the Discriminator of the DCGAN [10%]
The discriminator in this DCGAN is a convolutional neural network that has the following archi-tecture:
Discriminator
BatchNorm & ReLU
BatchNorm & ReLU
BatchNorm & ReLU
Sigmoid
32
16
8
4
1
8
4
1
16
1
32
64
128
32
conv2
conv3
conv4
conv1
3
Padding: In each of the convolutional layers shown above, we downsample the spatial di-mension of the input volume by a factor of 2. Given that we use kernel size K = 4 and stride S = 2, what should the padding be? Write your answer in your writeup, and show your work (e.g., the formula you used to derive the padding).
Implementation: Implement this architecture by lling in the __init__ method of the DCDiscriminator class in models.py, shown below. Note that the forward pass of DCDiscriminator is already provided for you.
def __init__(self, conv_dim=64):
super(DCDiscriminator, self).__init__()
###########################################
## FILL THIS IN: CREATE ARCHITECTURE ##
###########################################
self.conv1 = conv(...)
self.conv2 = conv(...)
self.conv3 = conv(...)
self.conv4 = conv(...)
Note: The function conv in models.py has an optional argument batch_norm: if batch_norm is False, then conv simply returns a torch.nn.Conv2d layer; if batch_norm is True, then conv returns a network block that consists of a Conv2d layer followed by a torch.nn.BatchNorm2d layer. Use the conv function in your implementation.
2
CSC321 Programming Assignment 4
Generator [10%]
Now, we will implement the generator of the DCGAN, which consists of a sequence of transpose convolutional layers that progressively upsample the input noise sample to generate a fake image. The generator we’ll use in this DCGAN has the following architecture:
Generator
BatchNorm & ReLU
BatchNorm & ReLU
BatchNorm & ReLU
tanh
32
1
4
8
16
100
1
4
8
16
128
64
32
deconv1
deconv2
deconv3
32
deconv4
3
Implementation: Implement this architecture by lling in the __init__ method of the DCGenerator class in models.py, shown below. Note that the forward pass of DCGenerator is already provided for you.
def __init__(self, noise_size, conv_dim):
super(DCGenerator, self).__init__()
###########################################
## FILL THIS IN: CREATE ARCHITECTURE ##
###########################################
self.deconv1 = deconv(...)
self.deconv2 = deconv(...)
self.deconv3 = deconv(...)
self.deconv4 = deconv(...)
Note: Use the deconv function (analogous to the conv function used for the discriminator above) in your generator implementation.
Training Loop [15%]
Next, you will implement the training loop for the DCGAN. A DCGAN is simply a GAN with a speci c type of generator and discriminator; thus, we train it in exactly the same way as a standard GAN. The pseudo-code for the training procedure is shown below. The actual implementation is simpler than it may seem from the pseudo-code: this will give you practice in translating math to code.
Implementation: Open up the le vanilla_gan.py and ll in the indicated parts of the training_loop function, starting at line 149, i.e., where it says
FILL THIS IN
1. Compute the discriminator loss on real images
D_real_loss = ...
3
CSC321 Programming Assignment 4
Algorithm 1 GAN Training Loop Pseudocode
procedure TrainGAN
Draw m training examples fx(1); : : : ; x(m)g from the data distribution pdata
Draw m noise samples fz(1); : : : ; z(m)g from the noise distribution pz
Generate fake images from the noise: G(z(i)) for i 2 f1; : : : :mg
Compute the (least-squares) discriminator loss:
J(D) =
1 m
D(x(i)) 1
2
+
1 m
D(G(z(i)))
2
2m i=1
2m i=1
X
X
Update the parameters of the discriminator
Draw m new noise samples fz(1); : : : ; z(m)g from the noise distribution pz
Generate fake images from the noise: G(z(i)) for i 2 f1; : : : :mg
Compute the (least-squares) generator loss:
J(G) =
1 m
D(G(z(i))) 1
2
m i=1
X
Update the parameters of the generator
There are 5 numbered bullets in the code to ll in for the discriminator and 3 bullets for the generator. Each of these can be done in a single line of code, although you will not lose marks for using multiple lines.
Experiment [10%]
1. Train the DCGAN with the command:
python vanilla_gan.py --num_epochs=40
By default, the script runs for 40 epochs (5680 iterations), and should take approximately 30 minutes on the teaching lab machines (it may be faster on your own computer). The script saves the output of the generator for a xed noise sample every 200 iterations throughout training; this allows you to see how the generator improves over time. Include in your write-up one of the samples from early in training (e.g., iteration 200) and one of the samples from later in training, and give the iteration number for those samples. Brie y comment on the quality of the samples, and in what way they improve through training.
Part 2: CycleGAN
Now we are going to implement the CycleGAN architecture.
Motivation: Image-to-Image Translation
Say you have a picture of a sunny landscape, and you wonder what it would look like in the rain. Or perhaps you wonder what a painter like Monet or van Gogh would see in it? These questions can be
4
CSC321 Programming Assignment 4
addressed through image-to-image translation wherein an input image is automatically converted into a new image with some desired appearance.
Recently, Generative Adversarial Networks have been successfully applied to image translation, and have sparked a resurgence of interest in the topic. The basic idea behind the GAN-based approaches is to use a conditional GAN to learn a mapping from input to output images. The loss functions of these approaches generally include extra terms (in addition to the standard GAN loss), to express constraints on the types of images that are generated.
A recently-introduced method for image-to-image translation called CycleGAN is particularly interesting because it allows us to use un-paired training data. This means that in order to train it to translate images from domain X to domain Y , we do not have to have exact correspondences between individual images in those domains. For example, in the paper that introduced CycleGANs, the authors are able to translate between images of horses and zebras, even though there are no images of a zebra in exactly the same position as a horse, and with exactly the same background, etc.
Thus, CycleGANs enable learning a mapping from one domain X (say, images of horses) to another domain Y (images of zebras) without having to nd perfectly matched training pairs.
To summarize the di erences between paired and un-paired data, we have:
Paired training data: f(x(i); y(i))gNi=1 Un-paired training data:
{ Source set: fx(i)gNi=1 with each x(i) 2 X
{ Target set: fy(j)gMj=1 with each y(j) 2 Y
{ For example, X is the set of horse pictures, and Y is the set of zebra pictures, where there are no direct correspondences between images in X and Y
Emoji CycleGAN
Now we’ll build a CycleGAN and use it to translate emojis between two di erent styles, in partic-ular, Windows $ Apple emojis.
Generator [20%]
The generator in the CycleGAN has layers that implement three stages of computation: 1) the rst stage encodes the input via a series of convolutional layers that extract the image features; 2) the second stage then transforms the features by passing them through one or more residual blocks; and 3) the third stage decodes the transformed features using a series of transpose convolutional layers, to build an output image of the same size as the input.
The residual block used in the transformation stage consists of a convolutional layer, where the input is added to the output of the convolution. This is done so that the characteristics of the output image (e.g., the shapes of objects) do not di er too much from the input.
Implement the following generator architecture by completing the __init__ method of the CycleGenerator class in models.py.
def __init__(self, conv_dim=64, init_zero_weights=False):
super(CycleGenerator, self).__init__()
###########################################
5
CSC321
Programming Assignment 4
GYtoXGXtoY
Apple
GXtoY
Windows
GYtoX
Apple
Emoji
Emoji
Emoji
conv
deconv
conv
deconv
conv
[0, 1]
Does the generated
image look like it came
DY
from the set of Windows
emojis?
## FILL THIS IN: CREATE ARCHITECTURE ##
###########################################
1. Define the encoder part of the generator
self.conv1 = ...
self.conv2 = ...
2. Define the transformation part of the generator
self.resnet_block = ...
3. Define the decoder part of the generator
self.deconv1 = ...
self.deconv2 = ...
To do this, you will need to use the conv and deconv functions, as well as the ResnetBlock class, all provided in models.py.
Note: There are two generators in the CycleGAN model, GX!Y and GY !X , but their imple-mentations are identical. Thus, in the code, GX!Y and GY !X are simply di erent instantiations of the same class.
6
CSC321
Programming Assignment 4
CycleGAN Generator
BatchNorm & ReLU
BatchNorm & ReLU
BatchNorm & ReLU
BatchNorm & ReLU
tanh
32
32
16
8
8
16
16
8
8
16
64
64
32
32
Redidual block
32
deconv1
32
conv1
conv2
deconv2
3
3
CycleGAN Training Loop [20%]
Finally, we will implement the CycleGAN training procedure, which is more involved than the procedure in Part 1.
Algorithm 2 CycleGAN Training Loop Pseudocode
procedure TrainCycleGAN
Draw a minibatch of samples fx(1); : : : ; x(m)g from domain X
Draw a minibatch of samples fy(1); : : : ; y(m)g from domain Y
Compute the discriminator loss on real images:
(D)
1
m
1
n
Xi
X
Jreal
= m
=1
(DX (x(i)) 1)2 + n
(DY (y(j) 1)2
j=1
Compute the discriminator loss on fake images:
(D)
1
m
1
n
Xi
X
Jfake
= m
=1
(DY (GX!Y (x(i))))2 + n
(DX (GY !X (y(j))))2
j=1
Update the discriminators
Compute the Y ! X generator loss:
1
n
J (GY !X) =
(DX (GY !X (y(j))) 1)2 + Jcycle(Y!X!Y )
n
Xj
=1
Compute the X ! Y generator loss:
1
m
J (GX!Y ) =
(DY (GX!Y (x(i))) 1)2 + Jcycle(X!Y !X)
m
Xi
=1
Update the generators
Similarly to Part 1, this training loop is not as di cult to implement as it may seem. There is a lot of symmetry in the training procedure, because all operations are done for both X ! Y and Y ! X directions. Complete the training_loop function in cycle_gan.py, starting from the following section:
# ============================================
# TRAIN THE DISCRIMINATORS
# ============================================
7
CSC321 Programming Assignment 4
#########################################
## FILL THIS IN ##
#########################################
Train with real images d_optimizer.zero_grad()
1. Compute the discriminator losses on real images
D_X_loss = ...
D_Y_loss = ...
There are 5 bullet points in the code for training the discriminators, and 6 bullet points in total for training the generators. Due to the symmetry between domains, several parts of the code you ll in will be identical except for swapping X and Y ; this is normal and expected.
Cycle Consistency
The most interesting idea behind CycleGANs (and the one from which they get their name) is the idea of introducing a cycle consistency loss to constrain the model. The idea is that when we translate an image from domain X to domain Y , and then translate the generated image back to domain X, the result should look like the original image that we started with.
The cycle consistency component of the loss is the mean squared error between the input images and their reconstructions obtained by passing through both generators in sequence (i.e., from domain X to Y via the X ! Y generator, and then from domain Y back to X via the Y ! X generator). The cycle consistency loss for the Y ! X ! Y cycle is expressed as follows:
m
X(y(i) GX!Y (GY !X (y(i))))2)
m
i=1
The loss for the X ! Y ! X cycle is analogous.
Implement the cycle consistency loss by lling in the following section in cycle_gan.py. Note that there are two such sections, and their implementations are identical except for swapping X and Y . You must implement both of them.
if opts.use_cycle_consistency_loss:
reconstructed_X = G_YtoX(fake_Y)
3. Compute the cycle consistency loss (the reconstruction loss)
cycle_consistency_loss = ...
g_loss += cycle_consistency_loss
CycleGAN Experiments [15%]
Training the CycleGAN from scratch can be time-consuming if you don’t have a GPU. In this part, you will train your models from scratch for just 600 iterations, to check the results. To save training time, we provide the weights of pre-trained models that you can load into your implementation. In order to load the weights, your implementation must be correct.
1. Train the CycleGAN without the cycle-consistency loss from scratch using the command:
8
CSC321 Programming Assignment 4
python cycle_gan.py
This runs for 600 iterations, and saves generated samples in the samples_cyclegan folder. In each sample, images from the source domain are shown with their translations to the right. Include in your writeup the samples from both generators at either iteration 400 or 600, e.g., sample-000400-X-Y.png and sample-000400-Y-X.png.
2. Train the CycleGAN with the cycle-consistency loss from scratch using the command:
python cycle_gan.py --use_cycle_consistency_loss
Similarly, this runs for 600 iterations, and saves generated samples in the samples_cyclegan_cycle folder. Include in your writeup the samples from both generators at either iteration 400 or 600 as above.
Now, we’ll switch to using pre-trained models, which have been trained for 40000 iterations. Run the pre-trained CycleGAN without the cycle-consistency loss using the command:
python cycle_gan.py --load=pretrained --train_iters=100
You only need 100 training iterations because the provided weights have already been trained
to convergence. The samples from the generators will be saved in the folder
samples_cyclegan_pretrained. Include the sampled output from your model.
4. Run the pre-trained CycleGAN with the cycle-consistency loss using the command:
python cycle_gan.py --load=pretrained_cycle \
--use_cycle_consistency_loss \
--train_iters=100
The samples will be saved in the folder samples_cyclegan_cycle_pretrained. Include the nal sampled output from your model.
Do you notice a di erence between the results with and without the cycle consistency loss? Write down your observations (positive or negative) in your writeup. Can you explain these results, i.e., why there is or isn’t a di erence between the two?
What you need to submit
Three code les: models.py, vanilla_gan.py, and cycle_gan.py.
A PDF document titled a4-writeup.pdf containing samples generated by your DCGAN and CycleGAN models, and your answers to the written questions.
9
CSC321 Programming Assignment 4
Further Resources
For further reading on GANs in general, and CycleGANs in particular, the following links may be useful:
Unpaired image-to-image translation using cycle-consistent adversarial networks (Zhu et al., 2017)
Generative Adversarial Nets (Goodfellow et al., 2014)
An Introduction to GANs in Tensor ow
Generative Models Blog Post from OpenAI
O cial PyTorch Implementations of Pix2Pix and CycleGAN
10
