<h1><big><center>MSO 3.4 - Deep Structured Learning</center></big></h1>
<h2><big><center> BE 2 - GANs and cGAN</center></big></h2>
<h2><big><center> BE 2 - GANs and Diffusion</center></big></h2>
<h5><big><center>Adapted from <i>Projet d'Option</i> of : Mhamed Jabri, Martin Chauvin, Ahmed Sahraoui, Zakariae Moustaïne and Taoufik Bouchikhi
<h5><center>GANs section adapted from the <i>Projet d'Option</i> of : Mhamed Jabri, Martin Chauvin, Ahmed Sahraoui, Zakariae Moustaïne and Taoufik Bouchikhi
<h5><center>Diffusion section implemented by Bruno Machado<br>
The aim of this assignment is to discover GANs, understand how they are implemented and then explore one specific architecture of GANs that allows us to perform image to image translation (which corresponds to the picture that you can see above this text ! )
The aim of this assignment is to discover generative models for images, understand how they are implemented and then explore some specific architectures and training strategies which allows us to perform image generation with and without conditioning (the picture above exemplifies image generation conditioned by another image)
Before starting the exploration of the world of GANs, here's what students should do and send back for this assignement :
* In the "tutorial" parts of this assignement that focus on explaining new concepts, you'll find <fontcolor='red'>**questions**</font> that aim to test your understanding of those concepts.
* In some of the code cells, you'll have to complete the code and you'll find a "TO DO" explaining what you should implement.
The deliverable for this assignment is this notebook, with the required modifications :
* In this notebook you'll find <fontcolor='red'>**questions**</font> that aim to test your understanding of the concepts studied. You should answer them by editing the Markdown cells (or adding new ones).
* In some of the code cells, you'll have to complete the code and you'll find a "TODO" explaining what you should implement.
* There are also some <fontcolor='green'>**bonus**</font> exercises. Those exercises are **optional** and should be performed only **AFTER** all the rest of the assignment is complete.
%% Cell type:markdown id: tags:
# Part1: DC-GAN
# Part1: DC-GAN
%% Cell type:markdown id: tags:
In this part, we aim to learn and understand the basic concepts of **Generative Adversarial Networks** through a DCGAN and generate new celebrities from the learned network after showing it real celebrities. For this purpose, please study the tutorial here: https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html
In this part, we aim to learn and understand the basic concepts of **Generative Adversarial Networks** through a DCGAN. For this purpose, please study the following tutorial: https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html
%% Cell type:markdown id: tags:
##Work to do
Now we want to generate handwritten digits using the MNIST dataset. It is available within torvision package (https://pytorch.org/vision/stable/generated/torchvision.datasets.MNIST.html#torchvision.datasets.MNIST)
Please re-train the DCGAN and display some automatically generated handwritten digits.
We want to generate handwritten digits using the MNIST dataset. It is available within the torvision package (https://pytorch.org/vision/stable/generated/torchvision.datasets.MNIST.html#torchvision.datasets.MNIST)
Adapt the code given in the tutorial in order to use it with the MNIST dataset. Present the loss curves in function of the gradient steps for the generator and discriminator and also compare the generated images with the images from the dataset.
%% Cell type:code id: tags:
``` python
#TODO: your code here to adapt the code from the tutorial to experiment on MNIST dataset
#TODO: adapt the code from the tutorial to experiment on the MNIST dataset
```
%% Cell type:markdown id: tags:
# Part2: Conditional GAN (cGAN)
The results for some images might be convincing, but probably you can spot some bad results as well. We will see how we can use different architectures and training objectives to get better results. More importantly, we generate images directly from noise, not knowing what number (if any) will come out on the output.
%% Cell type:markdown id: tags:
<fontcolor='red'>**Question**</font>
How could we change the architecture above to control which number the generator should output ?
%% Cell type:markdown id: tags:
Let's take the example of the set described in the next picture.
<fontcolor='green'>**Bonus**</font>
**Copy** the code you implemented above into a new Python cell and adapt it to implement your idea to control the number the generator outputs. Plot the training loss curves and a few examples of your model's outputs (e.g. generate(number=7) # Generates a handwritten 7). Note that we still want to be able to generate different images of the same class ! (i.e. generate(number=4) shoud not generate always the same handwritten 4, but a different one at each time)
%% Cell type:markdown id: tags:
# Part 2: Conditional GAN (cGAN)
%% Cell type:markdown id: tags:
Let's take the example of the set described in the following picture:
We have a picture of a map (from Google Maps) and we want to create an image of what the satellite view may look like.
As we are not only trying to generate a random picture but a mapping between a picture to another one, we can't use the standard GAN architecture. We will then use a cGAN introduced in this [paper](https://arxiv.org/pdf/1611.07004.pdf).
Therefore, we do not only want to generate a picture from random noise, but rather generate a picture from another picture. For this purpose we will use a cGAN instead of a vanilla GAN, which was introduced in this [paper](https://arxiv.org/pdf/1611.07004.pdf).
A cGAN is a supervised GAN aiming at mapping a label picture to a real one or a real picture to a label one. As you can see in the diagram below, the discriminator will take as input a pair of images and try to predict if the pair was generated or not. The generator will not only generate an image from noise but will also use an image (label or real) to generate another one (real or label).
A cGAN is a supervised GAN aiming at mapping a label picture to a real one or a real picture to a label one. As you can see in the diagram below, the discriminator will take as input a pair of images and try to predict if the pair was generated or not. The generator will not generate an image from noise but will intead use an image (label or real) to generate another one (real or label).
A U-Net takes as input an image, and outputs another image.
It can be divided into 2 subparts : an encoder and a decoder.
* The encoder takes the input image and reduces its dimension to encode the main features into a vector.
* The decoder takes this vector and map the features stored into an image.
A U-Net architecture is different from a classic encoder-decoder in that every layer of the decoder takes as input the previous decoded output as well as the output vector from the encoder layers of the same level. It allows the decoder to map low frequencies information encoded during the descent as well as high frequencies from the original picture.
A U-Net architecture is different from a vanilla convolutional encoder-decoder in that every layer of the decoder takes as input the previous decoded output as well as the output feature map from the encoder layers of the same level. This allows the decoder to map low frequencies information encoded during the descent as well as high frequencies from the original picture.
The encoder will take as input a colored picture (3 channels: RGB), it will pass through a series of convolution layers to encode the features of the picture. It will then be decoded by the decoder using transposed convolutional layers. These layers will take as input the previous decoded vector AND the encoded features of the same level.
%% Cell type:markdown id: tags:
Now, let's create or cGAN to generate facades from a template image. For this purpose, we will use the "Facade" dataset available at http://cmp.felk.cvut.cz/~tylecr1/facade/.
For this part the objective is to use a cGAN to generate facades from a template image. For this purpose, we will use the "Facade" dataset.
# At this stage x8 is our encoded vector, we will now decode it
x=self.up7(x8,x7)
x=self.up6(x,x6)
x=self.up5(x,x5)
x=self.up4(x,x4)
x=self.up3(x,x3)
x=self.up2(x,x2)
x=self.up1(x,x1)
x=self.outc(x)
returnx
```
%% Cell type:code id: tags:
``` python
# We take images that have 3 channels (RGB) as input and output an image that also have 3 channels (RGB)
generator=U_Net(3,3)
generator=U_Net(3,3)
# Check that the architecture is as expected
generator
print(generator)
```
%% Cell type:markdown id: tags:
You should now have a working U-Net.
%% Cell type:markdown id: tags:
<fontcolor='red'>**Question 1**</font>
Knowing the input and output images will be 256x256, what will be the dimension of the encoded vector x8 ?
<fontcolor='red'>**Question**</font>
Knowing the input and output images will have the shape 256x256 with 3 channels, what will be the dimension of the feature map "x8" ?
<fontcolor='red'>**Question 2**</font>
As you can see, U-net has an encoder-decoder architecture with skip connections. Explain why it works better than a traditional encoder-decoder.
<fontcolor='red'>**Question**</font>
As you can see, U-net has an encoder-decoder architecture with skip connections. Explain what is the point of the skip connections.
%% Cell type:markdown id: tags:
### Discriminator
In the cGAN architecture, the chosen discriminator is a Patch GAN. It is a convolutional discriminator which enables to produce a map of the input pictures where each pixel represents a patch of size NxN of the input.
In the cGAN architecture, the chosen discriminator is a Patch GAN. Instead of classifying if the whole image is fake or not (binary classification), this discriminator tries to classify if each N × N patch in an image is real or fake.
The size N is given by the depth of the net. According to this table :
| Number of layers | N |
| ---- | ---- |
| 1 | 16 |
| 2 | 34 |
| 3 | 70 |
| 4 | 142 |
| 5 | 286 |
| 6 | 574 |
The number of layers actually means the number of layers with `kernel=(4,4)`, `padding=(1,1)` and `stride=(2,2)`. These layers are followed by 2 layers with `kernel=(4,4)`, `padding=(1,1)` and `stride=(1,1)`.
In our case we are going to create a 70x70 PatchGAN.
As we want a 70x70 Patch GAN, the architecture will be as follows :
```
1. C64 - K4, P1, S2
2. C128 - K4, P1, S2
3. C256 - K4, P1, S2
4. C512 - K4, P1, S1
5. C1 - K4, P1, S1 (output)
```
Where Ck denotes a convolution block with k filters, Kk a kernel of size k, Pk is the padding size and Sk the stride applied.
*Note :* For the first layer, we do not use batchnorm.
%% Cell type:markdown id: tags:
<fontcolor='red'>**Question 3**</font>
Knowing input images will be 256x256 with 3 channels each, how many parameters are there to learn ?
<fontcolor='red'>**Question**</font>
Knowing input images will be 256x256 with 3 channels each, how many learnable parameters this neural network has ?
%% Cell type:code id: tags:
``` python
classPatchGAN(nn.Module):
def__init__(self,n_channels,n_classes):
super(PatchGAN,self).__init__()
# TODO :
# create the 4 first layers named conv1 to conv4
# TODO : create the 4 first layers named conv1 to conv4
self.conv1=
self.conv2=
self.conv3=
self.conv4=
# output layer
self.out=out_block(512,n_classes)
defforward(self,x1,x2):
x=torch.cat([x2,x1],dim=1)
x=self.conv1(x)
x=self.conv2(x)
x=self.conv3(x)
x=self.conv4(x)
x=self.out(x)
returnx
```
%% Cell type:code id: tags:
``` python
# We have 6 input channels as we concatenate 2 images (with 3 channels each)
discriminator=PatchGAN(6,1)
discriminator
discriminator=PatchGAN(6,1)
print(discriminator)
```
%% Cell type:markdown id: tags:
You should now have a working discriminator.
%% Cell type:markdown id: tags:
### Loss functions
As we have seen in the choice of the various architectures for this GAN, the issue is to map both low and high frequencies.
To tackle this problem, this GAN rely on the architecture to map the high frequencies (U-Net + PatchGAN) and the loss function to learn low frequencies features. The global loss function will indeed be made of 2 parts :
* the first part to map hight frequencies, will try to optimize the mean squared error of the GAN.
* the second part to map low frequencies, will minimize the $\mathcal{L}_1$ norm of the generated picture.
So the loss can be defined as $$ G^* = arg\ \underset{G}{min}\ \underset{D}{max}\ \mathcal{L}_{cGAN}(G,D) + \lambda \mathcal{L}_1(G)$$
The global loss function will be made of 2 parts :
* The vanilla GAN loss, in which the discriminator tries to maximize the probability it correctly classifies reals and fakes and the generator tries to minimize the probability that the discriminator will predict its outputs are fake.
* An auxiliary image reconstruction objective, in which the generator not only has to fool the discriminator, but also generate an image that is near to the ground truth image.
Therefore, the loss can be defined as: $$ G^* = arg\ \underset{G}{min}\ \underset{D}{max}\ \mathcal{L}_{cGAN}(G, D) + \lambda \mathcal{L}_{L1}(G)$$
print('There isn\' a training available with this number of epochs')
```
%% Cell type:code id: tags:
``` python
load_model(epoch=200)
# switching mode
generator.eval()
```
%% Cell type:code id: tags:
``` python
# show a sample evaluation image on the training base
image,mask=next(iter(dataloader))
output=generator(mask.type(Tensor))
output=output.view(16,3,256,256)
output=generator(mask.to(device))
output=output.cpu().detach()
foriinrange(8):
image_plot=reverse_transform(image[i])
output_plot=reverse_transform(output[i])
mask_plot=reverse_transform(mask[i])
plot2x3Array(mask_plot,image_plot,output_plot)
```
%% Cell type:code id: tags:
``` python
# show a sample evaluation image on the validation dataset
# show a sample evaluation image on the validation base
image,mask=next(iter(val_dataloader))
output=generator(mask.type(Tensor))
output=output.view(8,3,256,256)
output=generator(mask.to(device))
output=output.cpu().detach()
foriinrange(8):
image_plot=reverse_transform(image[i])
output_plot=reverse_transform(output[i])
mask_plot=reverse_transform(mask[i])
plot2x3Array(mask_plot,image_plot,output_plot)
```
%% Cell type:markdown id: tags:
<fontcolor='red'>**Question 4**</font>
Compare results for 100 and 200 epochs
<fontcolor='red'>**Question**</font>
Compare results of your model after 100 and 200 epochs
%% Cell type:code id: tags:
``` python
# TODO : Your code here to load and evaluate with a few samples the 2 checkpoints (100 epochs and 200 epochs)
```
%% Cell type:markdown id: tags:
# Part 3: Diffusion
%% Cell type:markdown id: tags:
Diffusion models are probabilistic generative models which learn to generate data by iteratively refining random noise through a reverse diffusion process. Given a sample of data, noise is progressvely added in small steps until it becomes pure noise. Then, a neural network is trained to reverse this process and generate realistic data from noise.
Diffusion models have gained popularity due to their ability to generate high-quality, diverse, and detailed content, surpassing GANs in the quality of the generated images.
In this assignment we will focus on DDPMs, which were introduced in this [paper](https://arxiv.org/abs/2006.11239) and laid the foundation for generative diffusion models.
%% Cell type:markdown id: tags:
For this part, we will use the MNIST dataset, used in part 1
%% Cell type:code id: tags:
``` python
# TODO: change the variable name to match the one you used in part 1 or reload the dataset
mnist_dataset=
mnist_dataloader=
```
%% Cell type:markdown id: tags:
Auxiliary function for plotting images
%% Cell type:code id: tags:
``` python
defplot1xNArray(images,labels):
f,axarr=plt.subplots(1,len(images))
forimage,ax,labelinzip(images,axarr,labels):
ax.imshow(image,cmap='gray')
ax.axis('off')
ax.set_title(label)
```
%% Cell type:markdown id: tags:
In order to train the model with the diffusion process, we will use a noise scheduler, which will be in charge of the forward diffusion process. The scheduler takes an image, a sample of random noise and a timestep, and return a noisy image for the corresponding timestep. Noise is progressvely added to the image at each timestep, therefore a noisy image at timestep 0 will have barely any noise while a noisy image at the maximum timestep will be basically just noise.
Let's create a noise scheduler with 1000 max timesteps and visualize some noise images.
We will use the diffusers library, which provides several tools for training and using diffusion models.
%% Cell type:code id: tags:
``` python
fromdiffusersimportDDPMScheduler
# TODO: Create the scheduler
noise_scheduler=
image,_=mnist_dataset[0]
# TODO: Create a noise tensor sampled from a normal distribution with the same shape as the image
For the reverse diffusion process we will use a neural network. Given a noisy image and the corresponding timestep, the goal of the neural network is to predict the noise, which allows for the denoising.
For the model, we will have a similar architecture as we used for the cGAN generator, a 2D UNet with a few modifications. The main difference will be that we have to indicate to the model which timestep is currently being denoised. For that purpose a timestep embedding is added, therefore the model has 2 inputs, the noisy image and the corresponding timestep.
In this exercise, we will use an UNet implementation from the diffusers library, which already has the timestep embedding included.
%% Cell type:code id: tags:
``` python
# TO DO : Your code here to load and evaluate with a few samples
What are the differences between the UNet used for the cGAN generator and the one defined above ?
Indicate the differences in the architecture by analyzing both models \_\_str\_\_.
%% Cell type:markdown id: tags:
<fontcolor='green'>**Bonus**</font>
**Extend** the code of the UNet used in part 2 to add the timestep conditioning. Train it in the same way as the UNet2DModel from the diffusers library (using the DDPMScheduler) and compare the results. As in the first bonus exercise, add a label conditioning which allows to control the output of the model (i.e. generate a specific handwritten digit (e.g. "4") instead of a random one). Test the model, showcasing some examples.
%% Cell type:markdown id: tags:
Now, let's train the model
%% Cell type:code id: tags:
``` python
# And finally :
ifcuda:
torch.cuda.empty_cache()
# ----------------
# Training Loop
# ----------------
torch.backends.cudnn.deterministic=True
losses=[]
num_epochs=5
print_every=100
diffusion_backbone.train()
forepochinrange(num_epochs):
fori,batchinenumerate(mnist_dataloader):
# Zero the gradients
optimizer.zero_grad()
# Send input to device
images=batch[0].to(device)
# Generate noisy images, different timestep for each image in the batch
If the training takes too long, you can download an already trained model from [this link](https://partage.liris.cnrs.fr/index.php/s/AP2t6b3w8SM4Bp5) and use it for inference.
%% Cell type:code id: tags:
``` python
# TODO: Add the path to the model checkpoint for loading the model
diffusion_backbone.load_state_dict(torch.load())
diffusion_backbone.eval()
```
%% Cell type:markdown id: tags:
Time to generate some images.
During training, for each data sample, we take a random timestep and correspondent noisy image to give it as input to our model. With suffitient training, the model should learn how to predict the noise in a noisy image for all possible timesteps.
During inference, to generate an image, we will start from pure noise and step by step predict the noise to go from one noisy image to the next, progressively denoising the image until we reach the timestep 0, in which we should have an image without any noise.
%% Cell type:code id: tags:
``` python
fromtqdmimporttqdm
# Start the image as random noise
image=torch.randn((10,1,64,64)).to(device)
# Create a list of images and labels for visualization
The diffusers library also provides *Pipeline* classes, which are wrappers around the model that abstracts the inference loop implemented above.
We can create a pipeline, giving it the trained model and the noise scheduler, and use it to generate images. In this case, we will only have access to the final image, generated on the last timestep, but not the intermediary images from the denoising process.
In this exercise we achieved decent results for very simple datasets. But we are quite far from those beautiful AI generated images we can find online. That is for 2 main reasons:
- Model size: due to the computation and time constrains, we can't really train very large models
- Dataset size: due to the same constrains, we can't use very complex and large datasets, which requires larger models and longer training times.
Fortunatly, even though we can train those large models with the available hardware and time, we can at least use them for inference !
The goal of this part is to learn how to retrive and use a pre-trained diffusion models and also to get creative to come up with some nice prompts to generate cool images.
We are going to use Stable Diffusion 3.5, which is a state of the art open-source text conditioned model. It takes a prompt in natural language and use it to guide the diffusion process. This type of models are trained with image-text pairs, but can generalize beyond the pairs seen during training, being able to mix several different concepts into a single image.
In order to save memory, we will use quantization, which consists into converting the model weights types from float16 into float4. That simply means that each model weight will be stored using only 4 bits instead of 16 bits. That allows us to run the model in GPUs with less VRAM and have faster inference, with a small drop in the quality of the results.
%% Cell type:markdown id: tags:
For this part of the assignment restart the notebook kernel to be sure your GPU memory is empty. The memory usage can be verified using the command `nvidia-smi` in a terminal. If your GPU has 2GB of VRAM or less, the model will probably not fit into memory even with quantization. In that case, use Google Colab for this part or use the smaller model indicated below. If you are not happy with the results and have plenty of VRAM available, feel free to increase the quantization to 8 bits or even load the model without quantization.
%% Cell type:markdown id: tags:
Before downloading the model, got to its [HuggingFace page](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium), log in with your account and accept the terms of use.
# TODO: test different prompts and visualize the generated images
# Once you are happy with the results, you can save 3 differet images as png file with the correspondent prompts in a text file
# Don't forget to add the images and prompts in your gitlab submission!
prompt=
image=pipeline(
prompt=prompt,
num_inference_steps=40,
guidance_scale=4.5,
max_sequence_length=512
).images[0]
image.save("generated_image.png")
```
%% Cell type:markdown id: tags:
If even with the quantization you still run out of GPU memory and you can't use Google Colab, you can use the following code instead, which uses a much smaller model (the results won't be as near as impressive, but it should be able to run even on a CPU, if you have a little patience)
This work must be done individually. The expected output is a private repository named gan-cgan on https://gitlab.ec-lyon.fr. It must contain your notebook (or python files) and a README.md file that explains briefly the successive steps of the project. Don't forget to add your teacher as developer member of the project. The last commit is due before 11:59 pm on Monday, April 1st, 2024. Subsequent commits will not be considered.
This work must be done individually. The expected output is a private repository named gan-diffusion on https://gitlab.ec-lyon.fr. It must contain your notebook (or python files), a README.md file that explains briefly the successive steps performed, the model checkpoints, and the generated images. Don't forget to add your teacher as developer member of the project. The last commit is due before 11:59 pm on Wednesday, April 9th, 2025. Subsequent commits will not be considered.
...
...
%% Cell type:markdown id: tags:
<h1><big><center>MSO 3.4 - Deep Structured Learning</center></big></h1>
<h2><big><center> BE 2 - GANs and cGAN</center></big></h2>
<h2><big><center> BE 2 - GANs and Diffusion</center></big></h2>
<h5><big><center>Adapted from <i>Projet d'Option</i> of : Mhamed Jabri, Martin Chauvin, Ahmed Sahraoui, Zakariae Moustaïne and Taoufik Bouchikhi
<h5><center>GANs section adapted from the <i>Projet d'Option</i> of : Mhamed Jabri, Martin Chauvin, Ahmed Sahraoui, Zakariae Moustaïne and Taoufik Bouchikhi
<h5><center>Diffusion section implemented by Bruno Machado<br>
The aim of this assignment is to discover GANs, understand how they are implemented and then explore one specific architecture of GANs that allows us to perform image to image translation (which corresponds to the picture that you can see above this text ! )
The aim of this assignment is to discover generative models for images, understand how they are implemented and then explore some specific architectures and training strategies which allows us to perform image generation with and without conditioning (the picture above exemplifies image generation conditioned by another image)
Before starting the exploration of the world of GANs, here's what students should do and send back for this assignement :
* In the "tutorial" parts of this assignement that focus on explaining new concepts, you'll find <fontcolor='red'>**questions**</font> that aim to test your understanding of those concepts.
* In some of the code cells, you'll have to complete the code and you'll find a "TO DO" explaining what you should implement.
The deliverable for this assignment is this notebook, with the required modifications :
* In this notebook you'll find <fontcolor='red'>**questions**</font> that aim to test your understanding of the concepts studied. You should answer them by editing the Markdown cells (or adding new ones).
* In some of the code cells, you'll have to complete the code and you'll find a "TODO" explaining what you should implement.
* There are also some <fontcolor='green'>**bonus**</font> exercises. Those exercises are **optional** and should be performed only **AFTER** all the rest of the assignment is complete.
%% Cell type:markdown id: tags:
# Part1: DC-GAN
# Part1: DC-GAN
%% Cell type:markdown id: tags:
In this part, we aim to learn and understand the basic concepts of **Generative Adversarial Networks** through a DCGAN and generate new celebrities from the learned network after showing it real celebrities. For this purpose, please study the tutorial here: https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html
In this part, we aim to learn and understand the basic concepts of **Generative Adversarial Networks** through a DCGAN. For this purpose, please study the following tutorial: https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html
%% Cell type:markdown id: tags:
##Work to do
Now we want to generate handwritten digits using the MNIST dataset. It is available within torvision package (https://pytorch.org/vision/stable/generated/torchvision.datasets.MNIST.html#torchvision.datasets.MNIST)
Please re-train the DCGAN and display some automatically generated handwritten digits.
We want to generate handwritten digits using the MNIST dataset. It is available within the torvision package (https://pytorch.org/vision/stable/generated/torchvision.datasets.MNIST.html#torchvision.datasets.MNIST)
Adapt the code given in the tutorial in order to use it with the MNIST dataset. Present the loss curves in function of the gradient steps for the generator and discriminator and also compare the generated images with the images from the dataset.
%% Cell type:code id: tags:
``` python
#TODO: your code here to adapt the code from the tutorial to experiment on MNIST dataset
#TODO: adapt the code from the tutorial to experiment on the MNIST dataset
```
%% Cell type:markdown id: tags:
# Part2: Conditional GAN (cGAN)
The results for some images might be convincing, but probably you can spot some bad results as well. We will see how we can use different architectures and training objectives to get better results. More importantly, we generate images directly from noise, not knowing what number (if any) will come out on the output.
%% Cell type:markdown id: tags:
<fontcolor='red'>**Question**</font>
How could we change the architecture above to control which number the generator should output ?
%% Cell type:markdown id: tags:
Let's take the example of the set described in the next picture.
<fontcolor='green'>**Bonus**</font>
**Copy** the code you implemented above into a new Python cell and adapt it to implement your idea to control the number the generator outputs. Plot the training loss curves and a few examples of your model's outputs (e.g. generate(number=7) # Generates a handwritten 7). Note that we still want to be able to generate different images of the same class ! (i.e. generate(number=4) shoud not generate always the same handwritten 4, but a different one at each time)
%% Cell type:markdown id: tags:
# Part 2: Conditional GAN (cGAN)
%% Cell type:markdown id: tags:
Let's take the example of the set described in the following picture:
We have a picture of a map (from Google Maps) and we want to create an image of what the satellite view may look like.
As we are not only trying to generate a random picture but a mapping between a picture to another one, we can't use the standard GAN architecture. We will then use a cGAN introduced in this [paper](https://arxiv.org/pdf/1611.07004.pdf).
Therefore, we do not only want to generate a picture from random noise, but rather generate a picture from another picture. For this purpose we will use a cGAN instead of a vanilla GAN, which was introduced in this [paper](https://arxiv.org/pdf/1611.07004.pdf).
A cGAN is a supervised GAN aiming at mapping a label picture to a real one or a real picture to a label one. As you can see in the diagram below, the discriminator will take as input a pair of images and try to predict if the pair was generated or not. The generator will not only generate an image from noise but will also use an image (label or real) to generate another one (real or label).
A cGAN is a supervised GAN aiming at mapping a label picture to a real one or a real picture to a label one. As you can see in the diagram below, the discriminator will take as input a pair of images and try to predict if the pair was generated or not. The generator will not generate an image from noise but will intead use an image (label or real) to generate another one (real or label).
A U-Net takes as input an image, and outputs another image.
It can be divided into 2 subparts : an encoder and a decoder.
* The encoder takes the input image and reduces its dimension to encode the main features into a vector.
* The decoder takes this vector and map the features stored into an image.
A U-Net architecture is different from a classic encoder-decoder in that every layer of the decoder takes as input the previous decoded output as well as the output vector from the encoder layers of the same level. It allows the decoder to map low frequencies information encoded during the descent as well as high frequencies from the original picture.
A U-Net architecture is different from a vanilla convolutional encoder-decoder in that every layer of the decoder takes as input the previous decoded output as well as the output feature map from the encoder layers of the same level. This allows the decoder to map low frequencies information encoded during the descent as well as high frequencies from the original picture.
The encoder will take as input a colored picture (3 channels: RGB), it will pass through a series of convolution layers to encode the features of the picture. It will then be decoded by the decoder using transposed convolutional layers. These layers will take as input the previous decoded vector AND the encoded features of the same level.
%% Cell type:markdown id: tags:
Now, let's create or cGAN to generate facades from a template image. For this purpose, we will use the "Facade" dataset available at http://cmp.felk.cvut.cz/~tylecr1/facade/.
For this part the objective is to use a cGAN to generate facades from a template image. For this purpose, we will use the "Facade" dataset.
# At this stage x8 is our encoded vector, we will now decode it
x=self.up7(x8,x7)
x=self.up6(x,x6)
x=self.up5(x,x5)
x=self.up4(x,x4)
x=self.up3(x,x3)
x=self.up2(x,x2)
x=self.up1(x,x1)
x=self.outc(x)
returnx
```
%% Cell type:code id: tags:
``` python
# We take images that have 3 channels (RGB) as input and output an image that also have 3 channels (RGB)
generator=U_Net(3,3)
generator=U_Net(3,3)
# Check that the architecture is as expected
generator
print(generator)
```
%% Cell type:markdown id: tags:
You should now have a working U-Net.
%% Cell type:markdown id: tags:
<fontcolor='red'>**Question 1**</font>
Knowing the input and output images will be 256x256, what will be the dimension of the encoded vector x8 ?
<fontcolor='red'>**Question**</font>
Knowing the input and output images will have the shape 256x256 with 3 channels, what will be the dimension of the feature map "x8" ?
<fontcolor='red'>**Question 2**</font>
As you can see, U-net has an encoder-decoder architecture with skip connections. Explain why it works better than a traditional encoder-decoder.
<fontcolor='red'>**Question**</font>
As you can see, U-net has an encoder-decoder architecture with skip connections. Explain what is the point of the skip connections.
%% Cell type:markdown id: tags:
### Discriminator
In the cGAN architecture, the chosen discriminator is a Patch GAN. It is a convolutional discriminator which enables to produce a map of the input pictures where each pixel represents a patch of size NxN of the input.
In the cGAN architecture, the chosen discriminator is a Patch GAN. Instead of classifying if the whole image is fake or not (binary classification), this discriminator tries to classify if each N × N patch in an image is real or fake.
The size N is given by the depth of the net. According to this table :
| Number of layers | N |
| ---- | ---- |
| 1 | 16 |
| 2 | 34 |
| 3 | 70 |
| 4 | 142 |
| 5 | 286 |
| 6 | 574 |
The number of layers actually means the number of layers with `kernel=(4,4)`, `padding=(1,1)` and `stride=(2,2)`. These layers are followed by 2 layers with `kernel=(4,4)`, `padding=(1,1)` and `stride=(1,1)`.
In our case we are going to create a 70x70 PatchGAN.
As we want a 70x70 Patch GAN, the architecture will be as follows :
```
1. C64 - K4, P1, S2
2. C128 - K4, P1, S2
3. C256 - K4, P1, S2
4. C512 - K4, P1, S1
5. C1 - K4, P1, S1 (output)
```
Where Ck denotes a convolution block with k filters, Kk a kernel of size k, Pk is the padding size and Sk the stride applied.
*Note :* For the first layer, we do not use batchnorm.
%% Cell type:markdown id: tags:
<fontcolor='red'>**Question 3**</font>
Knowing input images will be 256x256 with 3 channels each, how many parameters are there to learn ?
<fontcolor='red'>**Question**</font>
Knowing input images will be 256x256 with 3 channels each, how many learnable parameters this neural network has ?
%% Cell type:code id: tags:
``` python
classPatchGAN(nn.Module):
def__init__(self,n_channels,n_classes):
super(PatchGAN,self).__init__()
# TODO :
# create the 4 first layers named conv1 to conv4
# TODO : create the 4 first layers named conv1 to conv4
self.conv1=
self.conv2=
self.conv3=
self.conv4=
# output layer
self.out=out_block(512,n_classes)
defforward(self,x1,x2):
x=torch.cat([x2,x1],dim=1)
x=self.conv1(x)
x=self.conv2(x)
x=self.conv3(x)
x=self.conv4(x)
x=self.out(x)
returnx
```
%% Cell type:code id: tags:
``` python
# We have 6 input channels as we concatenate 2 images (with 3 channels each)
discriminator=PatchGAN(6,1)
discriminator
discriminator=PatchGAN(6,1)
print(discriminator)
```
%% Cell type:markdown id: tags:
You should now have a working discriminator.
%% Cell type:markdown id: tags:
### Loss functions
As we have seen in the choice of the various architectures for this GAN, the issue is to map both low and high frequencies.
To tackle this problem, this GAN rely on the architecture to map the high frequencies (U-Net + PatchGAN) and the loss function to learn low frequencies features. The global loss function will indeed be made of 2 parts :
* the first part to map hight frequencies, will try to optimize the mean squared error of the GAN.
* the second part to map low frequencies, will minimize the $\mathcal{L}_1$ norm of the generated picture.
So the loss can be defined as $$ G^* = arg\ \underset{G}{min}\ \underset{D}{max}\ \mathcal{L}_{cGAN}(G,D) + \lambda \mathcal{L}_1(G)$$
The global loss function will be made of 2 parts :
* The vanilla GAN loss, in which the discriminator tries to maximize the probability it correctly classifies reals and fakes and the generator tries to minimize the probability that the discriminator will predict its outputs are fake.
* An auxiliary image reconstruction objective, in which the generator not only has to fool the discriminator, but also generate an image that is near to the ground truth image.
Therefore, the loss can be defined as: $$ G^* = arg\ \underset{G}{min}\ \underset{D}{max}\ \mathcal{L}_{cGAN}(G, D) + \lambda \mathcal{L}_{L1}(G)$$
print('There isn\' a training available with this number of epochs')
```
%% Cell type:code id: tags:
``` python
load_model(epoch=200)
# switching mode
generator.eval()
```
%% Cell type:code id: tags:
``` python
# show a sample evaluation image on the training base
image,mask=next(iter(dataloader))
output=generator(mask.type(Tensor))
output=output.view(16,3,256,256)
output=generator(mask.to(device))
output=output.cpu().detach()
foriinrange(8):
image_plot=reverse_transform(image[i])
output_plot=reverse_transform(output[i])
mask_plot=reverse_transform(mask[i])
plot2x3Array(mask_plot,image_plot,output_plot)
```
%% Cell type:code id: tags:
``` python
# show a sample evaluation image on the validation dataset
# show a sample evaluation image on the validation base
image,mask=next(iter(val_dataloader))
output=generator(mask.type(Tensor))
output=output.view(8,3,256,256)
output=generator(mask.to(device))
output=output.cpu().detach()
foriinrange(8):
image_plot=reverse_transform(image[i])
output_plot=reverse_transform(output[i])
mask_plot=reverse_transform(mask[i])
plot2x3Array(mask_plot,image_plot,output_plot)
```
%% Cell type:markdown id: tags:
<fontcolor='red'>**Question 4**</font>
Compare results for 100 and 200 epochs
<fontcolor='red'>**Question**</font>
Compare results of your model after 100 and 200 epochs
%% Cell type:code id: tags:
``` python
# TODO : Your code here to load and evaluate with a few samples the 2 checkpoints (100 epochs and 200 epochs)
```
%% Cell type:markdown id: tags:
# Part 3: Diffusion
%% Cell type:markdown id: tags:
Diffusion models are probabilistic generative models which learn to generate data by iteratively refining random noise through a reverse diffusion process. Given a sample of data, noise is progressvely added in small steps until it becomes pure noise. Then, a neural network is trained to reverse this process and generate realistic data from noise.
Diffusion models have gained popularity due to their ability to generate high-quality, diverse, and detailed content, surpassing GANs in the quality of the generated images.
In this assignment we will focus on DDPMs, which were introduced in this [paper](https://arxiv.org/abs/2006.11239) and laid the foundation for generative diffusion models.
%% Cell type:markdown id: tags:
For this part, we will use the MNIST dataset, used in part 1
%% Cell type:code id: tags:
``` python
# TODO: change the variable name to match the one you used in part 1 or reload the dataset
mnist_dataset=
mnist_dataloader=
```
%% Cell type:markdown id: tags:
Auxiliary function for plotting images
%% Cell type:code id: tags:
``` python
defplot1xNArray(images,labels):
f,axarr=plt.subplots(1,len(images))
forimage,ax,labelinzip(images,axarr,labels):
ax.imshow(image,cmap='gray')
ax.axis('off')
ax.set_title(label)
```
%% Cell type:markdown id: tags:
In order to train the model with the diffusion process, we will use a noise scheduler, which will be in charge of the forward diffusion process. The scheduler takes an image, a sample of random noise and a timestep, and return a noisy image for the corresponding timestep. Noise is progressvely added to the image at each timestep, therefore a noisy image at timestep 0 will have barely any noise while a noisy image at the maximum timestep will be basically just noise.
Let's create a noise scheduler with 1000 max timesteps and visualize some noise images.
We will use the diffusers library, which provides several tools for training and using diffusion models.
%% Cell type:code id: tags:
``` python
fromdiffusersimportDDPMScheduler
# TODO: Create the scheduler
noise_scheduler=
image,_=mnist_dataset[0]
# TODO: Create a noise tensor sampled from a normal distribution with the same shape as the image
For the reverse diffusion process we will use a neural network. Given a noisy image and the corresponding timestep, the goal of the neural network is to predict the noise, which allows for the denoising.
For the model, we will have a similar architecture as we used for the cGAN generator, a 2D UNet with a few modifications. The main difference will be that we have to indicate to the model which timestep is currently being denoised. For that purpose a timestep embedding is added, therefore the model has 2 inputs, the noisy image and the corresponding timestep.
In this exercise, we will use an UNet implementation from the diffusers library, which already has the timestep embedding included.
%% Cell type:code id: tags:
``` python
# TO DO : Your code here to load and evaluate with a few samples
What are the differences between the UNet used for the cGAN generator and the one defined above ?
Indicate the differences in the architecture by analyzing both models \_\_str\_\_.
%% Cell type:markdown id: tags:
<fontcolor='green'>**Bonus**</font>
**Extend** the code of the UNet used in part 2 to add the timestep conditioning. Train it in the same way as the UNet2DModel from the diffusers library (using the DDPMScheduler) and compare the results. As in the first bonus exercise, add a label conditioning which allows to control the output of the model (i.e. generate a specific handwritten digit (e.g. "4") instead of a random one). Test the model, showcasing some examples.
%% Cell type:markdown id: tags:
Now, let's train the model
%% Cell type:code id: tags:
``` python
# And finally :
ifcuda:
torch.cuda.empty_cache()
# ----------------
# Training Loop
# ----------------
torch.backends.cudnn.deterministic=True
losses=[]
num_epochs=5
print_every=100
diffusion_backbone.train()
forepochinrange(num_epochs):
fori,batchinenumerate(mnist_dataloader):
# Zero the gradients
optimizer.zero_grad()
# Send input to device
images=batch[0].to(device)
# Generate noisy images, different timestep for each image in the batch
If the training takes too long, you can download an already trained model from [this link](https://partage.liris.cnrs.fr/index.php/s/AP2t6b3w8SM4Bp5) and use it for inference.
%% Cell type:code id: tags:
``` python
# TODO: Add the path to the model checkpoint for loading the model
diffusion_backbone.load_state_dict(torch.load())
diffusion_backbone.eval()
```
%% Cell type:markdown id: tags:
Time to generate some images.
During training, for each data sample, we take a random timestep and correspondent noisy image to give it as input to our model. With suffitient training, the model should learn how to predict the noise in a noisy image for all possible timesteps.
During inference, to generate an image, we will start from pure noise and step by step predict the noise to go from one noisy image to the next, progressively denoising the image until we reach the timestep 0, in which we should have an image without any noise.
%% Cell type:code id: tags:
``` python
fromtqdmimporttqdm
# Start the image as random noise
image=torch.randn((10,1,64,64)).to(device)
# Create a list of images and labels for visualization
The diffusers library also provides *Pipeline* classes, which are wrappers around the model that abstracts the inference loop implemented above.
We can create a pipeline, giving it the trained model and the noise scheduler, and use it to generate images. In this case, we will only have access to the final image, generated on the last timestep, but not the intermediary images from the denoising process.
In this exercise we achieved decent results for very simple datasets. But we are quite far from those beautiful AI generated images we can find online. That is for 2 main reasons:
- Model size: due to the computation and time constrains, we can't really train very large models
- Dataset size: due to the same constrains, we can't use very complex and large datasets, which requires larger models and longer training times.
Fortunatly, even though we can train those large models with the available hardware and time, we can at least use them for inference !
The goal of this part is to learn how to retrive and use a pre-trained diffusion models and also to get creative to come up with some nice prompts to generate cool images.
We are going to use Stable Diffusion 3.5, which is a state of the art open-source text conditioned model. It takes a prompt in natural language and use it to guide the diffusion process. This type of models are trained with image-text pairs, but can generalize beyond the pairs seen during training, being able to mix several different concepts into a single image.
In order to save memory, we will use quantization, which consists into converting the model weights types from float16 into float4. That simply means that each model weight will be stored using only 4 bits instead of 16 bits. That allows us to run the model in GPUs with less VRAM and have faster inference, with a small drop in the quality of the results.
%% Cell type:markdown id: tags:
For this part of the assignment restart the notebook kernel to be sure your GPU memory is empty. The memory usage can be verified using the command `nvidia-smi` in a terminal. If your GPU has 2GB of VRAM or less, the model will probably not fit into memory even with quantization. In that case, use Google Colab for this part or use the smaller model indicated below. If you are not happy with the results and have plenty of VRAM available, feel free to increase the quantization to 8 bits or even load the model without quantization.
%% Cell type:markdown id: tags:
Before downloading the model, got to its [HuggingFace page](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium), log in with your account and accept the terms of use.
# TODO: test different prompts and visualize the generated images
# Once you are happy with the results, you can save 3 differet images as png file with the correspondent prompts in a text file
# Don't forget to add the images and prompts in your gitlab submission!
prompt=
image=pipeline(
prompt=prompt,
num_inference_steps=40,
guidance_scale=4.5,
max_sequence_length=512
).images[0]
image.save("generated_image.png")
```
%% Cell type:markdown id: tags:
If even with the quantization you still run out of GPU memory and you can't use Google Colab, you can use the following code instead, which uses a much smaller model (the results won't be as near as impressive, but it should be able to run even on a CPU, if you have a little patience)
This work must be done individually. The expected output is a private repository named gan-cgan on https://gitlab.ec-lyon.fr. It must contain your notebook (or python files) and a README.md file that explains briefly the successive steps of the project. Don't forget to add your teacher as developer member of the project. The last commit is due before 11:59 pm on Monday, April 1st, 2024. Subsequent commits will not be considered.
This work must be done individually. The expected output is a private repository named gan-diffusion on https://gitlab.ec-lyon.fr. It must contain your notebook (or python files), a README.md file that explains briefly the successive steps performed, the model checkpoints, and the generated images. Don't forget to add your teacher as developer member of the project. The last commit is due before 11:59 pm on Wednesday, April 9th, 2025. Subsequent commits will not be considered.
We recommand to use the notebook (.ipynb) but the Python script (.py) is also provided if more convenient for you.
# How to submit your Work ?
This work must be done individually. The expected output is a private repository named gan-cgan on https://gitlab.ec-lyon.fr. It must contain your notebook (or python files) and a README.md file that explains briefly the successive steps of the project. Don't forget to add your teacher as developer member of the project. The last commit is due before 11:59 pm on Monday, April 1st, 2024. Subsequent commits will not be considered.
This work must be done individually. The expected output is a private repository named gan-diffusion on https://gitlab.ec-lyon.fr. It must contain your notebook (or python files) and a README.md file that explains briefly the successive steps of the project. Don't forget to add your teacher as developer member of the project. The last commit is due before 11:59 pm on Wednesday, April 9th, 2025. Subsequent commits will not be considered.
Dellandrea Emmanuel authored