Hello, I have got a torchgan base with version v0.0.4. Which version of pytorch should be used? I did not find any documentation on this.

best,
Sixin

List of GAN Models we want to have before v0.1

• DCGAN
• Conditional GAN
• Autoencoder Model for EBGAN and BEGAN
• Self Attention GAN

@avik-pal Add any other models you find relevant.

The code doesn't respect the way we accept the directory name for the checkpoints. The name of the directory is the part of the str appearing before the /. Should be trivial to fix.

Describe the bug
The process to resume training a previously trained model is unintuitive.

To Reproduce
Steps to reproduce the behavior:

1. Package Versions: torchgan 0.1.0, torchvision 0.13.1, pytorch 1.12.1
2. Logging Configurations:

   print(torchgan.logging.backends.CONSOLE_LOGGING)
   1
   print(torchgan.logging.backends.VISDOM_LOGGING)
   0
   print(torchgan.logging.backends.TENSORBOARD_LOGGING)
   0

3. Minimal Working Example for the error
   The issue can be easily encountered by slightly modifying Tutorial 1. Follow the tutorial normally, until you reach the "Visualizing the Samples" section. Before that section, add the following code cell:

   trainer.load_model("./model/gan4.model")
   trainer(dataloader)

   Now execute the new cell.

Expected behavior
This should have continued training the model for an additional 10 generations on CUDA, or 5 generations otherwise. However, because the Trainer's epochs parameter represents the total number of epochs to train, the function returns without any further training done. In order to achieve the expected behaviour, the user must instead create a new Trainer object and pass (current_epochs + desired_additional_epochs) as the value of the epochs parameter. This is unintuitive, and requires the user to manually keep track of how many epochs have been completed when they end a training session if they plan on continuing it later.

Desktop (please complete the following information):

• OS: Windows 10 Pro, version 21H2

Installation

• Pip

Additional context
Fixing this would involve rewriting significant portions of the BaseTrainer class. I would suggest allowing the user to pass in a number of epochs to train in the __call__() function, rather than having it set at object creation.

Currently, instead of displaying the gradients norm we are showing the norm of the model parameters.
Problems in fixing this simply:

1. We are clearing the gradient buffers every time; this makes it impossible to track gradients for different losses.
2. Printing the gradient wrt only the last executed loss will lead to wrong inference on the part of the user.
3. Can't make gradients a property of the loss as it makes custom losses a nightmare to implement (we will have to impose a lot of constraints).

I cannot find any reference to TTUR in the documentation (see more about the technique here https://arxiv.org/abs/1706.08500). Would it be possible to implement it for WGAN-GP for torchgan? I'm new to the framework but I think the right place to implement it would be in torchgan.losses.WassersteinGradientPenalty.

I've been using TorchGAN to train the global generator/discriminator of pix2pixHD and got great results. I wonder if it's supported to load the pre-trained global models, add local enhancer layers on top of it, and continue the training?

If the current trainer doesn't support this, could I have some directions on how to implement it? Thanks!

I noticed that minimax_discriminator_loss uses binary_cross_entropy_with_logits (cf. binary_cross_entropy) which "combines a Sigmoid layer and the BCELoss" according to the docs.

This is a good practice when it comes to numerical stability, but maybe emphasizing this in the docs would be a good idea to prevent users from inserting a sigmoid as the final non-linearity in custom models. Maybe even raising a warning if we can somehow detect this.

The loss-ops and sampling from the noise priors are too entangled with one another, and the entire library is written keeping a Normal Distribution in mind for the noise prior. This leads to inconveniences such as overriding the entire train ops of a loss (and even the sampler) just to change something as minor as changing the noise prior from Normal to Uniform.

I propose having a separate set of modules for priors, keeping both fixed and learnable priors in mind. Will submit a PR regarding the same.

Describe the bug
An error will be reported if the 0.2.0 version of fastprogress is installed

Traceback (most recent call last):
  File "dc_gan.py", line 205, in <module>
    trainer(dataloader)
  File "/usr/local/anaconda3/envs/pt1.0py37/lib/python3.7/site-packages/torchgan/trainer/base_trainer.py", line 455, in __call__
    self.train(data_loader, **kwargs)
  File "/usr/local/anaconda3/envs/pt1.0py37/lib/python3.7/site-packages/torchgan/trainer/base_trainer.py", line 400, in train
    master_bar_iter.first_bar.comment = f"Training Progress"
AttributeError: 'ConsoleMasterBar' object has no attribute 'first_bar'

I used the 0.2.0 version fastprogress to create a simple example

from fastprogress import master_bar, progress_bar
master_bar_iter = master_bar(range(0, 3))
master_bar_iter.first_bar.comment = f'first bar stat'

Produced a similar error

AttributeError                            Traceback (most recent call last)
<ipython-input-3-ffd0b33d372e> in <module>
----> 1 master_bar_iter.first_bar.comment = f'first bar stat'

AttributeError: 'ConsoleMasterBar' object has no attribute 'first_bar'

I won't get an error when I install 0.1.20 version fastprogress

pip install fastprogress==0.1.20

RuntimeError: output with shape [1, 32, 32] doesn't match the broadcast shape [3, 32, 32].

While trying to display images from dataloader on Google Colab.

List of Loss Functions we want to have before v0.1 arranged roughly in order of priority

• Minimax Discriminator Loss
• Minimax Generator Loss
• Minimax Discriminator Loss with Nonsaturating Heuristic
• Minimax Generator Loss with Nonsaturating Heuristic
• Wasserstein Discriminator Loss
• Wasserstein Generator Loss
• Wasserstein Gradient Penalty
• Least Squares Discriminator Loss
• Least Squares Generator Loss
• Mutual Information Penalty
• EBGAN Discriminator Loss
• EBGAN Generator Loss
• ACGAN Discriminator Loss
• ACGAN Generator Loss
• EBGAN Repelling Regularizer
• DRAGAN Discriminator Loss

@Aniket1998 add more losses you find relevant and needed for initial release.

This needs to be done in 2 parts:

• Separate out the functional forms of the losses into a module of losses
• Document the necessary functions

We don't need to expose all the functions. Some functions do nothing fancy and they need to be removed and the entire computation can be performed inside the forward function.

Describe the bug
Some confusion regarding the implementation of virtual batch norm.

1. Referring to another code: https://github.com/dansuh17/segan-pytorch/blob/938b6c837b6f091d0ad1e67fad9db32045a90151/vbnorm.py#L68 , it seems that in the torchgan version the mean and deviation are set to None after a batch has passed. in the next batch these will be computed for that sample and applied, and not from a "virtual batch".

   torchgan/torchgan/layers/virtualbatchnorm.py

   Line 93 in f413953

   self.ref_var = None

   . so it seems for an alternate batch it uses statistics from that batch, and for the rest from the previous batch.
2. even when the statistics are updated they are directly copied from a batch as opposed to interpolated as in https://github.com/dansuh17/segan-pytorch/blob/938b6c837b6f091d0ad1e67fad9db32045a90151/vbnorm.py#L66

Traceback (most recent call last):
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/site-packages/urllib3/connectionpool.py", line 600, in urlopen
chunked=chunked)
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/site-packages/urllib3/connectionpool.py", line 354, in _make_request
conn.request(method, url, **httplib_request_kw)
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/http/client.py", line 1239, in request
self._send_request(method, url, body, headers, encode_chunked)
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/http/client.py", line 1285, in _send_request
self.endheaders(body, encode_chunked=encode_chunked)
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/http/client.py", line 1234, in endheaders
self._send_output(message_body, encode_chunked=encode_chunked)
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/http/client.py", line 1026, in _send_output
self.send(msg)
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/http/client.py", line 964, in send
self.connect()
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/site-packages/urllib3/connection.py", line 181, in connect
conn = self._new_conn()
File "/home/movan/anaconda3/envs/tensorflow/lib/python3.6/site-packages/urllib3/connection.py", line 168, in _new_conn
self, "Failed to establish a new connection: %s" % e)
urllib3.exceptions.NewConnectionError: <urllib3.connection.HTTPConnection object at 0x7f75df0ba6a0>: Failed to establish a new connection: [Errno 111] Connection refused

When I run the AAE tutorial, the output fake image is as following. For each image, only left top part is clear. Is there any problem related to this outcome?

Issue Description

I am reviewing the JOSS submission linked to this repository, see: Joss Review

Running

python -m pytest --cov=torchgan /tests --cov-report term-missing

showcases that the test coverage for the losses module for the repository is quite low. Are you planning on expanding your loss test cases? I think this is important for this type of package as it drives the model training, and is such linked to the checkbox "functionality" for the review.

I implemented RGAN and RCGAN from [1]. Would it make sense to add it to torchgan.models?

[1] Esteban, Cristóbal, Stephanie L. Hyland, and Gunnar Rätsch. "Real-valued (medical) time series generation with recurrent conditional gans." arXiv preprint arXiv:1706.02633 (2017).

The exact approach to sampling images from the trainer at the end of every epoch depends on what the generator requires as input during its forward pass. The current sample_images method is mainly geared towards generators that accept noise input (with a small workaround for generators that require labels). This needs to be generalised properly and be made more flexible to allow users to have greater control on the sampling process.

In CycleGAN Tutorial.ipynb, the description and code about identity loss are wrong.
Current code is :
fake_a = gen_b2a(image_b)
fake_b = gen_a2b(image_a)
loss_identity = 0.5 * (F.l1_loss(fake_a, image_a) + F.l1_loss(fake_b, image_b))

The correct code should be:
loss_identity = 0.5 * (F.l1_loss(gen_b2a(image_a), image_a) + F.l1_loss(gen_a2b(image_b), image_b))

Describe the bug
Getting the following error when trying to train on images of size larger than 32x32:
"RuntimeError: shape '[64]' is invalid for input of size 1600"

To Reproduce
My hyperparameters for the networks as well as the data I am loading are below

tfs = transforms.Compose([
    transforms.Resize(64),
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
train_dataset = torchvision.datasets.ImageFolder(root=data_dir+"train", transform=tfs)
val_dataset = torchvision.datasets.ImageFolder(root=data_dir+"/val", transform=tfs)
test_dataset = torchvision.datasets.ImageFolder(root=data_dir+"/test", transform=tfs)

dataset = train_dataset

dataloader = torch.utils.data.DataLoader(dataset, batch_size=64, shuffle=True, num_workers=4)


dcgan_network = {
    "generator": {
        "name": DCGANGenerator,
        "args": {
            "encoding_dims": 100,
            "out_channels": 3,
            "step_channels": 64,
            "nonlinearity": nn.LeakyReLU(0.2),
            "last_nonlinearity": nn.Tanh(),
        },
        "optimizer": {"name": Adam, "args": {"lr": 0.0001, "betas": (0.5, 0.999)}},
    },
    "discriminator": {
        "name": DCGANDiscriminator,
        "args": {
            "in_channels": 3,
            "step_channels": 64,
            "nonlinearity": nn.LeakyReLU(0.2),
            "last_nonlinearity": nn.LeakyReLU(0.2),
        },
        "optimizer": {"name": Adam, "args": {"lr": 0.0003, "betas": (0.5, 0.999)}},
    },
}```

**Expected behavior**
Normal training as usual

**Installation**
- Conda

The current framework is too much oriented towards Computer Vision problems where there is already lots of available resources. It would be great if you can also focus on multivariate time series where Generative Adversarial Networks can have several very important uses in data generation, augmentation, imputation, denoising and detecting anomalies for several applications such as medical, financial, etc.. It would be great if you can add examples for few of these use-cases. Enclosed is such multi-variate time series dataset consisting of log-returns where cubic interpolation is employed for imputation and market with a Boolean, I am sending it in case you would like to have an example for such a common use-case in financial world.

Dataset.zip

Describe the bug
torchgan install tests to site-packages folder, not site-packages/torchgan
To Reproduce
Steps to reproduce the behavior:

1. torchgan 0.0.2.r33.g8c2bc99

Expected behavior
A clear and concise description of what you expected to happen.

Desktop (please complete the following information):

• OS: ArchLinux

Installation

• Build from source, see build() and package() function in PKGBUILD.

Additional context
You might consider moving tests folder to torchgan folder.

The following evaluation metrics need to be implemented

• Inception Score
• Frechet Distance
• Classifier Score measured with a custom user-trained classifier
• Sliced Wasserstein Distance

@avik-pal Add any more metrics that you think, should be part of the list

I've been using this project the last couple of weeks to run the experiments for a paper I'm writing.
I was wondering how exactly I could cite it.

https://help.github.com/en/articles/referencing-and-citing-content

The time duration of running of epoch of epoch should be printed in epoch summary as it happens in tensorflow . Please tell which files i would need to change , I would like to work on this.

When I run the code in torchgan/tutorials/Introduction To TorchGAN.ipynb, I got this error!
How can I fix it?

This line is showing error when torchgan is run locally . Should I change it to
from .models import Discriminator, Generator
as in other files like acgan.py

The Logger implemented in #50 was just to get the things running. It is not very user-friendly making it very difficult to use custom Visualizers, even extending the logger to new Visualizers by making modifications in the core library.

#50 attempts to add support for 2 backends. Also, we drop any dependency on plotting and visualization library through this PR. In future, we would like to support other plotting backends.
Some of the prioritised backends are:-

• Tensorboard
• Visdom
• Matplotlib
• Seaborn
• Bokeh

However, this is not a high priority issue and will not be dealt with till a stable version has been released.

Following models are proposed to be added:

• Adversarial Autoencoders
• Adversarial Variational Bayes
• ALI/BiGAN

Since we are pushing for the release and putting more focus on feature completeness, I am listing all the current modules. All of these need to be documented before the next release can be tagged.

• Loss Functions: New functions lack the equations. We also don't have a dedicated documentation for the functional forms of the losses. However, the train_ops need to have proper documentation as it is necessary for customizability.
• Models: Models are fairly well documented. They just need some minor clean ups.
• Metrics: The metrics API is a bit difficult to use. If we can't find a cleaner way to deal with metrics the only way is to have well-documented examples.
• Logger: Lacks Documentation
• Layers: Lacks Documentation
• Trainer: Documentation is fine. However, needs a few demonstrative examples.

These need to be numeric tests and ideally tested with implementations across other frameworks for consistency checking.

We need to modify the _get_argument_maps() function to handle this. We should check if the argument is optional and avoid throwing error in that case.
This should be pretty easy to solve. We just need to pass everything as a keyword argument instead of a positional argument.

Hi, you have a mistake in the conditions at row 75 and 167 in torchgan/layers/residual.py.

You want if i != len(filters) - 1 :)

Figured it out!

The code is being documented along with the PRs. However, we need to portray some good examples and also use latex formatting wherever needed.
This will be done once we roll out v0.1-alpha.
Hence not a high priority task. However, individual methods need to be documented from the start.

Is your feature request related to a problem? Please describe.
I successfully use DCGAN to generate avatars trained by cartoonset (https://google.github.io/cartoonset/). I want to make an improvement that allows specifying attributes such as hair color (among 10 choices) and face shape (among 7 choices), not just randomly generating an avatar.

Describe the solution you'd like
Could you please add one GAN network to support multiple attributes?

Describe alternatives you've considered
I tried CGAN but it seems that it only supports one attribute at a time.

Running example/self-attention GAN error
Traceback (most recent call last):
File "/home1/code/test_gan/self-att-GAN/self-attGAN.py", line 121, in
trainer(dataloader)
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torchgan-0.0.2-py3.6.egg/torchgan/trainer/trainer.py", line 454, in call
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torchgan-0.0.2-py3.6.egg/torchgan/trainer/trainer.py", line 416, in train
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torchgan-0.0.2-py3.6.egg/torchgan/trainer/trainer.py", line 343, in train_iter
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torchgan-0.0.2-py3.6.egg/torchgan/losses/loss.py", line 76, in train_ops
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in call
result = self.forward(*input, **kwargs)
File "/home1/code/test_gan/self-att-GAN/self-attGAN.py", line 42, in forward
return self.model(x)
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in call
result = self.forward(*input, **kwargs)
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/container.py", line 92, in forward
input = module(input)
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in call
result = self.forward(*input, **kwargs)
File "/home/xujin/anaconda3/lib/python3.6/site-packages/torchgan-0.0.2-py3.6.egg/torchgan/layers/spectralnorm.py", line 73, in forward
RuntimeError: cublas runtime error : the GPU program failed to execute at /pytorch/aten/src/THC/THCBlas.cu:116

The coverage results I have tested locally are around 70% but in travis they seem to be reported as low as 22%. Apparently none od the files inside torchgan are being covered.

If you have used torchgan in your publication (even if it is only a preprint version as of now), open a pull request to add it to the README, or simply post it here (tagging @avik-pal). We will add it to the README file.

I am getting the error AttributeError: 'VirtualBatchNorm' object has no attribute 'weight' while using the module to implement virtual batch normalization. For the context, here is my weights_init function:
`

def weights_init(m):
classname = m.class.name
print(classname)
if classname.find('Conv') != -1:
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif classname.find('BatchNorm') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0)`
python version -3.7.4
pyTorch version-1.5.0

Can torchgan be installed for win10 ?
Are there ways to install torchgan in win10 ? I installed anaconda 3-4.2.0 with python3.5, tensorflow, torch and torchvision.

dist = torch.distributions.one_hot_categorical.OneHotCategorical(torch.tensor([0.1]*10))

lossf = MutualInformationPenalty()
z = torch.randn(2, 100)
c_cont = torch.randn(2, 30)
c_dis = dist.sample([2])
print(c_dis)

G = InfoGANGenerator(10, 30)
D = InfoGANDiscriminator(10, 30)

fake = G(z, c_dis, c_cont)

score, q_categorical, q_gaussian = D(fake, return_latents=True)

loss = lossf(c_dis, c_cont, q_categorical, q_gaussian)
loss.backward()

> Traceback (most recent call last):
  File "test.py", line 40, in <module>
    loss.backward()
  File "/usr/local/lib/python3.5/dist-packages/torch/tensor.py", line 102, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File "/usr/local/lib/python3.5/dist-packages/torch/autograd/__init__.py", line 90, in backward
    allow_unreachable=True)  # allow_unreachable flag
RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

Hi,

There are quite a few errors in the documentation. And it took me a lot of time to build the model correctly.

When I used the above code, there were some errors and I was wondering if it was my usage error.

Thanks.

Background

TorchGAN started out with a primary focus on GANs. But recent literature has suggested that GANs are being used in time series modeling, NLP, etc. Hence the name Trainer sends the wrong message.

Proposal

1. Deprecate Trainer and ParallelTrainer in v0.0.4 and completely remove them in v0.0.5+
2. Introduce an ImageTrainer and ParallelImageTrainer which are exact copies of the current version of Trainer and ParallelTrainer.
3. Introduce application-specific Trainers to broaden the scope of TorchGAN. #110 discusses this.

When training is resumed using load_model, there is no change in the state of the Logger. So if we start to retrain at epoch x, the images and all step value will be stored starting again from 1.

I implemented it exactly as in the example/ example, but it is particularly poor compared to the example display.Can I ignore any key details?

• Historical Averaging
• Energy Based GAN (EBGAN-PT cannot be confirmed before the basic losses start giving proper results) [ Most likely the error is due to hyperparameters]

Thanks a lot for this project! Is there any possibility to add some models which are designed for generating HD images i.e. StyleGAN (StyleGAN2, StyleGAN2-ADA)?

Thanks in advance!