akanimax / msg-gan-v1 Goto Github PK

View Code? Open in Web Editor NEW

151.0 10.0 28.0 203.59 MB

MSG-GAN: Multi-Scale Gradients GAN (Architecture inspired from ProGAN but doesn't use layer-wise growing)

License: MIT License

Python 100.00%

msg-gan gan machine-learning deep-learning adversarial-machine-learning celeba

msg-gan-v1's Introduction

**Please note that this is not the repo for the MSG-GAN research paper. Please head over to the msg-stylegan-tf repository for the official code and trained models for the MSG-GAN paper.

MSG-GAN

MSG-GAN (Multi-Scale Gradients GAN): A Network architecture inspired from the ProGAN.

The architecture of this gan contains connections between the intermediate layers of the singular Generator and the Discriminator. The network is not trained by progressively growing the layers. All the layers get trained at the same time.

Implementation uses the PyTorch framework.

Celeba samples

Please note that all the samples at various scales are generated by the network simultaneously.

Multi-Scale Gradients architecture

The above figure describes the architecture of the proposed Multi-Scale gradients GAN. As you can notice, from every intermediate layer of the Generator, a particular resolution image is extracted through (1 x 1) convolutions. These extracted images are in turn fed to the appropriate layers of the Discriminator. This allows for gradients to flow from the Discriminator to the Generator at multiple scales.

For the discrimination process, appropriately downsampled versions of the real images are fed to corresponding layers of the discriminator as shown in the diagram.

The problem of occurence of random gradients for GANs at the higher resolutions is tackled by layerwise training in the ProGAN paper. I present another solution for it. I have run the following experiment that preliminarily validates the proposed approach.

Above figure explains how the Meaningful Gradients penetrate the Generator from Bottoms-up. Initially, only the lower resolution gradients are menaingful and thus start generating good images at those resolutions, but eventually, all the scales synchronize and start producing images. This results in a stabler training for the higher resolution.

Celeba Experiment

I ran the experiment on a skimmed version of the architecture as described in the ProGAN paper. Following table summarize the details of the Networks:

For extracting images after every 3 layer block at that resolution, I used 1 x 1 convolutions. Similar operation is performed for feeding the images to discriminator intermediate layers.

The architecture for the discriminator is also the same (reverse mirror), with the distinction that half of the channels come from the (1 x 1 convolution) transformed downsampled real images and half from conventional top-to-bottom path.

All the 3 x 3 convolution weights have a forward hook that applies spectral normalization on them. Apart from that, in the discriminator for the 4 x 4 layer, there is a MinibatchStd layer for improving sample diversity. No other stablization techniques are applied.

64 x 64 experiment

128 x 128 experiment

The above diagrams are the loss plots obtained during training the Networks in an adversarial manner. The loss function used is Relativistic Hinge-GAN. Apart from some initial aberrations, the training has stayed smooth.

Running the Code

Please note to use value of learning_rate=0.0003 for both G and D for all experiments. TTUR doesn't work with this architecture (from experience). And, you can find other better learning rates, but the value 0.0003 always seems to work.

Running the training is actually very simple. Just start the training by running the train.py script in the sourcecode/ directory. Refer to the following parameters for tweaking for your own use:

-h, --help            show this help message and exit
 --generator_file GENERATOR_FILE
                    pretrained weights file for generator
 --discriminator_file DISCRIMINATOR_FILE
                    pretrained_weights file for discriminator
 --images_dir IMAGES_DIR
                    path for the images directory
 --sample_dir SAMPLE_DIR
                    path for the generated samples directory
 --model_dir MODEL_DIR
                    path for saved models directory
 --loss_function LOSS_FUNCTION
                    loss function to be used: 'hinge', 'relativistic-
                    hinge'
 --depth DEPTH         Depth of the GAN
 --latent_size LATENT_SIZE
                    latent size for the generator
 --batch_size BATCH_SIZE
                    batch_size for training
 --start START         starting epoch number
 --num_epochs NUM_EPOCHS
                    number of epochs for training
 --feedback_factor FEEDBACK_FACTOR
                    number of logs to generate per epoch
 --num_samples NUM_SAMPLES
                    number of samples to generate for creating the grid
                    should be a square number preferably
 --gen_dilation GEN_DILATION
                    amount of dilation for the generator
 --dis_dilation DIS_DILATION
                    amount of dilation for the discriminator
 --checkpoint_factor CHECKPOINT_FACTOR
                    save model per n epochs
 --g_lr G_LR           learning rate for generator
 --d_lr D_LR           learning rate for discriminator
 --adam_beta1 ADAM_BETA1
                    value of beta_1 for adam optimizer
 --adam_beta2 ADAM_BETA2
                    value of beta_2 for adam optimizer
 --use_spectral_norm USE_SPECTRAL_NORM
                    Whether to use spectral normalization or not
 --data_percentage DATA_PERCENTAGE
                    percentage of data to use
 --num_workers NUM_WORKERS
                    number of parallel workers for reading files

Running 1024 x 1024 architecture

For training a network as per the ProGAN CelebaHQ experiment, use the following arguments:

$ python train.py --depth=9 \
                  --latent_size=512 \
                  --images_dir=<path to CelebaHQ images> \
                  --sample_dir=samples/CelebaHQ_experiment \
                  --model_dir=models/CelebaHQ_experiment

Set the batch_size, feedback_factor and checkpoint_factor accordingly. This experiment was carried out by me on a DGX-1 machine. The samples displayed in Figure 1. of this readme are the output of this experiment. You can use the models pretrained for 3 epochs at [1024 x 1024] for your training. These are available at -> https://drive.google.com/drive/folders/119n0CoMDGq2K1dnnGpOA3gOf4RwFAGFs

Trained weights for generating cool faces :)

Please refer to the models/Celeba/1/GAN_GEN_3.pth for the saved weights for this model in PyTorch format.

Thanks

Please feel free to open PRs here if you train on other datasets using this architecture.

Best regards,
@akanimax :)

msg-gan-v1's People

Contributors

Stargazers

Watchers

msg-gan-v1's Issues

about Paper <MSG-GAN>

Hi, I am reading your paper arxiv : 1903.06048v1 .
In your paper the github address is BMSG-GAN but I did not find it in your repos.

How to connect intermediate layer of generator to discriminator ??

about MSG-GAN paper

Hi, I look your paper and code of this repo.
This code is not implement :

PixNorm,
Equalized Learning Rate ,
Exponential Moving Averaging of Generator Parameters
??? Is it true ?

Working with Gray scale images

Inorder to work with medical gray images, Pls let me know how the code has to be modified.
I find in_channels & out_channels in GAN.py. Is that to be modified? or in CustomLayers.py?

No Need to compute Generator Gradients when training Discriminator

Since Discriminator and Generator are trained separately, we don't need the gradients of the generator when training the discriminator.

Here are some relevant issues:
https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/d7e7ed851b3e5ac01167197011bf8204015a1ce7/models/cycle_gan_model.py#L140
pytorch/pytorch#5349

Fixing this will probably speed up training.

CUDA OOM During Training Due to self.create_grid().

For whatever reason, self.create_grid(self.gen(fixed_input), gen_img_files) seems to use a massive amount of GPU memory when training, so much so that I cannot train on 4TitanX with 1 image per GPU. There is probably some substantial optimization that can be done here to reduce the memory overhead. After removing it, I went from having out of memory errors to have 7GB freed on each GPU.

No FC layers?

ProGAN, and most GANs, use some fully-connected/dense layers in between the latent z vector and the first CNN/upscaling layers in the Generator. Most strikingly, StyleGAN uses no less than 8 FC layers as part of the 'style mapping' network which is then the input into its Generator.

The graph in the README implies that there's some FC layers (I think; the colors are so similar I can't tell where the FC layers are supposed to be), but they don't seem to be in the code? Looking at the CustomLayers.py's GenInitialBlock, it seems to be pure convolutions, and the printed model arch also seems to not have. Adding them is easy:

diff --git a/sourcecode/MSG_GAN/CustomLayers.py b/sourcecode/MSG_GAN/CustomLayers.py
index 381d9f8..bc91929 100644
--- a/sourcecode/MSG_GAN/CustomLayers.py
+++ b/sourcecode/MSG_GAN/CustomLayers.py
@@ -19,9 +19,13 @@ class GenInitialBlock(th.nn.Module):
         :param in_channels: number of input channels to the block
         """
         from torch.nn import LeakyReLU
-        from torch.nn import Conv2d, ConvTranspose2d
+        from torch.nn import Conv2d, ConvTranspose2d, Linear
         super().__init__()
 
+        # from torch.nn import LeakyReLU
+        self.fc1 = th.nn.Linear(in_channels, in_channels)
+        self.fc2 = th.nn.Linear(in_channels, in_channels)
+
         self.conv_1 = ConvTranspose2d(in_channels, in_channels, (4, 4), bias=True)
         self.conv_2 = Conv2d(in_channels, in_channels, (3, 3), padding=(1, 1), bias=True)
 
@@ -34,10 +38,14 @@ class GenInitialBlock(th.nn.Module):
         :param x: input to the module
         :return: y => output
         """
+        x = self.lrelu(self.fc1(x))
+        x = self.lrelu(self.fc2(x))
+
         # convert the tensor shape:
         y = x.view(*x.shape, 1, 1)  # add two dummy dimensions for
         # convolution operation

Seems to help training of some anime faces noticeably over the past few hours - although I also added self-attention layers to the D/G so not a clean comparison. Some overnight samples: https://imgur.com/a/CbnaWB0

Since most GANs do that, ProGAN does it and is the mentioned inspiration, and the graph seems to imply you intended to do it, maybe you want to add some fcs somewhere.

Specify format of CelebA-HQ

I was going to try to use some GPUs to get results for CelebA-HQ dataset in it's numpy format doesn't seem to work with this repo. What format do you expect CelebA-HQ to be in?

I want to just test the discriminator of this GAN using my abnormal images

Hi, @Skylion007 @Ianmcmill @akanimax

I trained successfully anomalyGAN with your MSG-GAN using my only normal images.

And now , I want to just test the discriminator of this GAN using my abnormal images...

How can I extract the discriminator from your MSG-GAN ?

Thanks in advance.

Best,
@bemoregt.

ask for a pretrained model 128*128

Hi, I am very interested in your work. Could you please provide the pretrained model on size 128*128?
Thank you very much.

Difference between this repo and BMSG-GAN

Hello,

It's really nice repo.

But could you please specify the difference betweent your another repo BMSG-GAN and this repo?

Seems this repo is somehow older. But what is the exactly update in BMSG-GAN?

thanks and best regards.

WGAN-GP loss with multiple images

Hi,
Thanks for creating this repository! I'm implementing the MSG-GAN method for my own project and I'm trying to use WGAN-GP loss, but I'm finding it difficult since there are multiple images. Could you point to the WGAN-GP loss implementation with multiple images in this repo so that I can use it for reference?
Thanks,
Tharun Mohandoss

Sizes of tensors must match except in dimension 0. Got 3 and 1 in dimension 1

I am not sure what is the important error.
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 3 and 1 in dimension 1 at c:\new-builder_3\win-wheel\pytorch\aten\src\th\generic/THTensorMath.cpp:3616

RuntimeError: Couldn't open shared event: <torch_5308_2995310656_event>, error code: <2>

I converted the images to RGB with
mogrify -colorspace RGB *.jpg
but that didn't help.
Any hints?

Starting the training process ...

Epoch: 1
Elapsed [0:00:08.859118] batch: 1  d_loss: 2.000845  g_loss: 3.779521
Elapsed [0:00:10.685150] batch: 3  d_loss: 0.985300  g_loss: 4.276993
Elapsed [0:00:13.235160] batch: 6  d_loss: 1.903335  g_loss: 3.318109
Elapsed [0:00:15.852519] batch: 9  d_loss: 1.755461  g_loss: 3.607772
Traceback (most recent call last):
  File "train.py", line 221, in <module>
    main(parse_arguments())
  File "train.py", line 215, in main
    start=args.start
  File "F:\portrait\MSG-GAN\sourcecode\MSG_GAN\GAN.py", line 423, in train
    for (i, batch) in enumerate(data, 1):
  File "C:\Users\castle\Envs\keras_tut\lib\site-packages\torch\utils\data\dataloader.py", line 336, in __next__
    return self._process_next_batch(batch)
  File "C:\Users\castle\Envs\keras_tut\lib\site-packages\torch\utils\data\dataloader.py", line 357, in _process_next_batch
    raise batch.exc_type(batch.exc_msg)
RuntimeError: Traceback (most recent call last):
  File "C:\Users\castle\Envs\keras_tut\lib\site-packages\torch\utils\data\dataloader.py", line 106, in _worker_loop
    samples = collate_fn([dataset[i] for i in batch_indices])
  File "C:\Users\castle\Envs\keras_tut\lib\site-packages\torch\utils\data\dataloader.py", line 164, in default_collate
    return torch.stack(batch, 0, out=out)
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 3 and 1 in dimension 1 at c:\new-builder_3\win-wheel\pytorch\aten\src\th\generic/THTensorMath.cpp:3616

Exception ignored in: <bound method _DataLoaderIter.__del__ of <torch.utils.data.dataloader._DataLoaderIter object at 0x000002C3EC473978>>
Traceback (most recent call last):
  File "C:\Users\castle\Envs\keras_tut\lib\site-packages\torch\utils\data\dataloader.py", line 399, in __del__
    self._shutdown_workers()
  File "C:\Users\castle\Envs\keras_tut\lib\site-packages\torch\utils\data\dataloader.py", line 378, in _shutdown_workers
    self.worker_result_queue.get()
  File "c:\users\castle\appdata\local\programs\python\python36\Lib\multiprocessing\queues.py", line 337, in get
    return _ForkingPickler.loads(res)
  File "C:\Users\castle\Envs\keras_tut\lib\site-packages\torch\multiprocessing\reductions.py", line 167, in rebuild_storage_filename
    storage = cls._new_shared_filename(manager, handle, size)
RuntimeError: Couldn't open shared event: <torch_5308_2995310656_event>, error code: <2>