rosinality / alias-free-gan-pytorch Goto Github PK

There is a check that doesn't allow to use conv2d_gradfix with PyTorch 1.9.0 (or later):

I tried running this on multiple GPUs and noticed all the processes allocate memory on GPU number 1. This essentially limits me to a batch size of just 1 when each GPU can support closer to a batch size of 4 on their own.

Dear Rosinality,

Thank you for you great implementation.

Would you mind providing the pretrained models? Maybe for unaligned FFHQ?

Thank you for your help.

Best Wishes,

Alex

Hey @rosinality exciting repo!

I'm working on a fork with pytorch lightning for training on tpus but I've hit a roadblock where it's having trouble loading images. So I thought I should try running your training script to make sure it was something I changed about the dataloader.

Switched to a fresh install of your repo to run the test.

Using a colab pro instance with a Tesla P100-PCIE-16GB.

Downloaded a couple pip libraries to get things working (tensorfn, wandb, ninja, and jsonnet). Converted my dataset. Then changed the config file use a size of 256. And ran the following command:

!python train.py --n_gpu 1 --conf /content/drive/MyDrive/afg-lightning-devel/alias-free-gan-pytorch/config/config-t-256.jsonnet training.batch=16 path="/content/drive/MyDrive/afg-lightning-devel/alias-free-gan-pytorch/datasets/painterly-faces-256"

The memory error I'm getting:

Output appended
...
  0% 0/800000 [00:00<?, ?it/s]/content/drive/MyDrive/afg-lightning-devel/alias-free-gan-pytorch/stylegan2/op/conv2d_gradfix.py:89: UserWarning: conv2d_gradfix not supported on PyTorch 1.9.0+cu102. Falling back to torch.nn.functional.conv2d().
  f"conv2d_gradfix not supported on PyTorch {torch.__version__}. Falling back to torch.nn.functional.conv2d()."
Traceback (most recent call last):
  File "train.py", line 406, in <module>
    main, conf.n_gpu, conf.n_machine, conf.machine_rank, conf.dist_url, args=(conf,)
  File "/usr/local/lib/python3.7/dist-packages/tensorfn/distributed/launch.py", line 49, in launch
    fn(*args)
  File "train.py", line 399, in main
    train(conf, loader, generator, discriminator, g_optim, d_optim, g_ema, device)
  File "train.py", line 250, in train
    fake_img = generator(noise)
  File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/content/drive/MyDrive/afg-lightning-devel/alias-free-gan-pytorch/model.py", line 424, in forward
    out = conv(out, latent)
  File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/content/drive/MyDrive/afg-lightning-devel/alias-free-gan-pytorch/model.py", line 303, in forward
    out = self.activation(out)
  File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/content/drive/MyDrive/afg-lightning-devel/alias-free-gan-pytorch/model.py", line 258, in forward
    out = fused_leaky_relu(out, negative_slope=self.negative_slope)
  File "/content/drive/MyDrive/afg-lightning-devel/alias-free-gan-pytorch/stylegan2/op/fused_act.py", line 119, in fused_leaky_relu
    return FusedLeakyReLUFunction.apply(input, bias, negative_slope, scale)
  File "/content/drive/MyDrive/afg-lightning-devel/alias-free-gan-pytorch/stylegan2/op/fused_act.py", line 66, in forward
    out = fused.fused_bias_act(input, bias, empty, 3, 0, negative_slope, scale)
RuntimeError: CUDA out of memory. Tried to allocate 1.16 GiB (GPU 0; 15.90 GiB total capacity; 11.59 GiB already allocated; 231.75 MiB free; 14.75 GiB reserved in total by PyTorch)
  0% 0/800000 [00:11<?, ?it/s]

How much memory does your config require? Do I need to decrease my batch size (or other settings) to be able to train on colab?

Hi, sorry for asking a dumb question. I had a quite hard time (but still failed) to understand the function of filter_parameters in https://github.com/rosinality/alias-free-gan-pytorch/blob/main/model.py#L47

It would be grateful if you can add some comments or explanations to this part.
Moreover, especially if we want to change the input dimensions (instead of 16x16 for now), how should we modify this function?

Thanks

Hello, thanks for your great work.

I tried to train alias-free gan with config-t on the resolution of 1024, by simply editing training.size to 1024 in the config file, but the network seems not converge and the generated result is not good either.

Do I need to modify some other arguments for training on a larger resolution (e.g. 512, 1024)?

Thanks!

Hi. I wonder why you use just the jinc(||x||), whereas in the original article they use jinc(2fc||x||).

Traceback (most recent call last):
File "train.py", line 404, in
conf = load_arg_config(GANConfig)
File "/home/user/anaconda3/envs/pytorch/lib/python3.8/site-packages/tensorfn/util/config.py", line 269, in load_arg_config
conf = load_config(config_model, args.conf, args.opts, show)
File "/home/user/anaconda3/envs/pytorch/lib/python3.8/site-packages/tensorfn/util/config.py", line 257, in load_config
conf = config_model(**read_config(config, overrides=overrides))
File "/home/user/anaconda3/envs/pytorch/lib/python3.8/site-packages/tensorfn/util/config.py", line 25, in read_config
json_str = _jsonnet.evaluate_file(config_file)
AttributeError: 'NoneType' object has no attribute 'evaluate_file'

data format.

-Data
--Images
---1.jpg
|
|
|
---N.jpg

Hi there, thanks so much for this implementation! I know this is probably trivial for more advanced users but could you provide some tips or code to accelerate my ability to generate images and movies. Something like what you did for StyleGAN2/ADA would be amazing.

please add colab demo

Can anybody provide a setting to train 1024 img? The existing setting seems only support 256 img.

Hi!
The authors from the paper said that they needed to use a custom conda kernel to get acceptable performance (They say it's 20x faster than with pytorch primitives. I noticed u didn't use a custom cuda object. How is your performance compared to StyleGAN2?

Thanks!

nice job.
how to project real images and interpolate them in alias free gan.

Hi, I'm trying to find out which cutoff frequencies you actually used for you experiments.

So, I tried to find where is the function 'filter_parameters' is being used, but I can't seem to find it.

Could you point me to where it is being set?

Thanks!

Hi!
First of all, thank you for sharing the implementation!

I was wondering about how you have generated the sample videos in the README file. Is there a code in the repo? In the project page from NVIDIA, they stated that the videos are generated by randomly walking around a central point in the latent space.

Thank you.

Thanks for the great contribution.

Slightly off-topic, but do you know how they generate the internal representations in Figure 6 in the original paper? ie. what convention do they use to visualise the channels in the feature maps as RGB?

Thanks for this repo, it's great!

To get it working in colab, I copied the bare minimum out from the docker file:

!pip install jsonnet
!apt install -y -q ninja-build
!pip install tensorfn rich

!pip install setuptools
!pip install numpy scipy nltk lmdb cython pydantic pyhocon

!apt install libsm6 libxext6 libxrender1
!pip install opencv-python-headless

It then works despite throwing two compatibility errors:

ERROR: requests 2.23.0 has requirement urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1, but you'll have urllib3 1.26.6 which is incompatible.
ERROR: datascience 0.10.6 has requirement folium==0.2.1, but you'll have folium 0.8.3 which is incompatible.

I then made some manual edits to config/config-t.jsonnet so it runs on colab:

Under training:{} set image size to 128
Under training:{} batch size to 12 (650mb each so <8gb I guess)

In prepare_data.py I commented out line 14 for no resizing, just cropping. Could be useful config for some datasets.

In train.py main function line 322 and comment out 5 "logger" lines. the logger info didn't work, it just hangs then falls over without error out of the box in colab but I didn't investigate further.

I also couldn't get --ckpt=checkpoint/010000.pt to resume properly. I tried editing start iteration in the config too but no luck, it just seemed to start from zero again.

Also, it may be worth editing train.py with autocast() for half precision float16 instead of float32 to improve speed and memory limitations? Or even porting to TPU? https://github.com/pytorch/xla

So then run

!git clone https://github.com/rosinality/alias-free-gan-pytorch.git

After making these edits

#upload your zip file or use google drive import
!unzip /content/dataraw.zip -d /content/dataraw

%cd /content/alias-free-gan-pytorch
!python prepare_data.py --out /content/dataset --n_worker 8 --size=128 /content/dataraw

%cd /content/alias-free-gan-pytorch
!python train.py --n_gpu 1 --conf config/config-t.jsonnet path=/content/dataset/

Thanks again!

README.md says "This implementation contains a lot of my guesses, so I think there are many differences to the official implementations". Is this sentence still accurate?

Hi, thanks for the work. I have a question regarding the input fourier features.
I think you might have included the margin into the target canvas (-0.5~0.5) ?
That makes the input frequencies become relatively lower.

A possible fix would be something like:

class FourierFeature(nn.Module):
    def __init__(self, size=16, margin=10, dim=512, cutoff=2, eps=1e-8):
        """
        size:   sampling rate (or feature map size)
        margin: expanded feature map margin size
        dim:    # channels
        cutoff: cutoff fc
        """
        super().__init__()

        normalized_margin = margin / size
        # -0.5-m ~ 0.5+m, uniform interplate 'size' (except the last one)
        # note the margin here, target canvas was -0.5~0.5, extended canvas should be larger
        coords = torch.linspace(- 0.5 - normalized_margin, 
                                  0.5 + normalized_margin, 
                                size + 2 * margin + 1)[:-1]

Hi! I wanted to try this GAN, but I don't have enough memory to run it. I have a 1080ti 11gb and used training.batch=1
Is there any way to optimize the network so that it fit?

I´ve been trying to run the repo via colab. Everything goes fine until running the train script, then the error below is thrown. Has the repo been recently updated or am I missing some dependency or some other detail to be considered? Any clue how to solve this?

IndexError Traceback (most recent call last) /usr/local/lib/python3.7/dist-packages/pyparsing.py in _parseNoCache(self, instring, loc, doActions, callPreParse) 1682 try: -> 1683 loc, tokens = self.parseImpl(instring, preloc, doActions) 1684 except IndexError: 8 frames IndexError: string index out of range During handling of the above exception, another exception occurred: ParseException Traceback (most recent call last) ParseException: Expected {: ... | {{{"=" | ":" | "+="} - [Suppress:({{{"#" | "//"} - SkipTo:({Suppress:(W:( )) | StringEnd})} | Suppress:(W:( ))})]... -} ConcatenatedValueParser:([{{{{Suppress:({[Suppress:(W:( ,))] {"#" | "//"} - SkipTo:({Suppress:(W:( )) | StringEnd})}) | {Suppress:("include") {Re:('"(?:[^"\\\n]|\\.)"[ \t]') | {{"url" | "file" | "package"} - Suppress:("(") - Re:('"(?:[^"\\\n]|\\.)"[ \t]') - Suppress:(")")} | {"required" - Suppress:("(") - {Re:('"(?:[^"\\\n]|\\.)"[ \t]') | {{"url" | "file" | "package"} - Suppress:("(") - Re:('"(?:[^"\\\n]|\\.)"[ \t]') - Suppress:(")")}} - Suppress:(")")}}} | Re:('[ \t]\$\{[^\}]+\}[ \t]') | : ...} | Forward: {{{{Suppress:("[") -} ListParser:({{ConcatenatedValueParser:([{{{{Suppress:({[Suppress:(W:( ,))] {"#" | "//"} - SkipTo:({Suppress:(W:( )) | StringEnd})}) | {Suppress:("include") {Re:('"(?:[^"\\\n]|\\.)"[ \t]') | {{"url" | "file" | "package"} - Suppress:("(") - Re:('"(?:[^"\\\n]|\\.)"[ \t]') - Suppress:(")")} | {"required" - Suppress:("(") - {Re:('"(?:[^"\\\n]|\\.)"[ \t]') | {{"url" | "file" | "package"} - Suppress:("(") - Re:('"(?:[^"\\\n]|\\.)"[ \t]') - Suppress:(")")}} - Suppress:(")")}}} | Re:('[ \t]\$\{[^\}]+\}[ \t]') | : ...} | : ...} | {{{{{{{{W:(0123...) Suppress:([]...)} {"ns" ^ "nano" ^ "nanos" ^ "nanosecond" ^ "nanoseconds" ^ "us" ^ "micro" ^ "micros" ^ "microsecond" ^ "microseconds" ^ "ms" ^ "milli" ^ "millis" ^ "millisecond" ^ "milliseconds" ^ "s" ^ "second" ^ "seconds" ^ "m" ^ "minute" ^ "minutes" ^ "h" ^ "hour" ^ "hours" ^ "w" ^ "week" ^ "weeks" ^ "d" ^ "day" ^ "days" ^ "mo" ^ "month" ^ "months" ^ "y" ^ "year" ^ "years"}} Supp} | {{{{{{{{W:(0123...) Suppress:([]...)} {"ns" ^ "nano" ^ "nanos" ^ "nanosecond" ^ "nanoseconds" ^ "us" ^ "micro" ^ "micros" ^ "microsecond" ^ "microseconds" ^ "ms" ^ "milli" ^ "millis" ^ "millisecond" ^ "milliseconds" ^ "s" ^ "second" ^ "seconds" ^ "m" ^ "minute" ^ "minutes" ^ "h" ^ "hour" ^ "hours" ^ "w" ^ "week" ^ "weeks" ^ "d" ^ "day" ^ "days" ^ "mo" ^ "month" ^ "months" ^ "y" ^ "year" ^ "years"}} Suppress:(WordEnd)} | Re:('[+-]?(\d*\.\d+|\d+(\.\d+)?)([eE][+\-]?\d+)?(?=$|[ \t]([\$\}\],#\n\r]|//))')} | "true"} | "false"} | "null"} | {{Re:('""".?""""') | Re:('"(?:[^"\\\n]|\\.)"[ \t]*')} | Re:('(... ))})}]...)}}, found end of text (at char 43), (line:1, col:44)

I have been experimenting with how to add a couple of extra values for scaling the Fourier features. I still think my implementation is wrong because what should look like a smooth zoom-in looks more like if we were changing the FoV of the camera, so things in the center of the image scale at a slightly different rate. Some tweaks need to be done for this to work seamlessly, but the results are still pretty cool. With some refinement, would you consider adding this as a configurable parameter? (once i manage to make it work as I want, I could open a PR)

@rosinality thank you for always putting out such fantastic work. I have a questions about your training details: How much GPU memory is required, at minimum, to run this implementation of alias free gan? In the paper, the authors mention training on "n NVIDIA DGX-1 with 8 Tesla V100 GPUs", but no mention of how much GPU memory is required. Also, how long did it take you to train?

When I set augment to true in the config file I get a RuntimeError on lines 66 and 49 of stylegan2/op/fused_act.py

RuntimeError: input must be contiguous

Unless I make the input contiguous.

gradgrad_out = fused.fused_bias_act( gradgrad_input.contiguous(), gradgrad_bias, out, 3, 1, ctx.negative_slope, ctx.scale )

out = fused.fused_bias_act(input.contiguous(), bias, empty, 3, 0, negative_slope, scale)

train is too slow

I'm getting the following error in stylegan2/op/fused_act.py at line 66:

RuntimeError: CUDA error: an illegal memory access was encountered

Did not use the docker file and I'm running on torch 1.8.1 and I'm running on a 3090

Hi. I wonder what are those artifacts on the generated images?

It seems like they appear on your title animation with faces, too.