Implementation of 'lightweight' GAN, proposed in ICLR 2021, in Pytorch. High resolution image generations that can be trained within a day or two

@lucidrains

After update to this version https://github.com/lucidrains/lightweight-gan/releases/tag/0.15.0 I cant continue train my network and did start in from zero. Previous version was in state 117k batches by 4 (468k images, around 66 hours of trainig) image and was pretty good. In new version 0.15.0 on same dataset with same parameters (--image-size 1024 --aug-types [color,offset_h] --aug-prob 1 --amp --batch-size 7) after 77k batches by 7 (539k images, around 49 hours of training) I see some bugs like oil puddle. Did you meet this or do you know how avoid this?

In previous version with sle-spatial I didnt meet something like this.

I have seen this function argument mentioned in this issue:

https://github.com/lucidrains/lightweight-gan/issues/14#issuecomment-733432989

What is sle_spatial?

I have the following problem since today. How to do/solve this?

continuing from previous epoch - 118 loading from version 0.21.4 unable to load save model. please try downgrading the package to the version specified by the saved model Traceback (most recent call last): File "/opt/conda/bin/lightweight_gan", line 8, in sys.exit(main()) File "/opt/conda/lib/python3.8/site-packages/lightweight_gan/cli.py", line 193, in main fire.Fire(train_from_folder) File "/opt/conda/lib/python3.8/site-packages/fire/core.py", line 141, in Fire component_trace = _Fire(component, args, parsed_flag_args, context, name) File "/opt/conda/lib/python3.8/site-packages/fire/core.py", line 466, in _Fire component, remaining_args = _CallAndUpdateTrace( File "/opt/conda/lib/python3.8/site-packages/fire/core.py", line 681, in _CallAndUpdateTrace component = fn(*varargs, **kwargs) File "/opt/conda/lib/python3.8/site-packages/lightweight_gan/cli.py", line 184, in train_from_folder run_training(0, 1, model_args, data, load_from, new, num_train_steps, name, seed, use_aim, aim_repo, aim_run_hash) File "/opt/conda/lib/python3.8/site-packages/lightweight_gan/cli.py", line 59, in run_training model.load(load_from) File "/opt/conda/lib/python3.8/site-packages/lightweight_gan/lightweight_gan.py", line 1603, in load raise e File "/opt/conda/lib/python3.8/site-packages/lightweight_gan/lightweight_gan.py", line 1600, in load self.GAN.load_state_dict(load_data['GAN']) File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1497, in load_state_dict raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state_dict for LightweightGAN: Missing key(s) in state_dict: "G.layers.0.0.2.1.weight", "G.layers.0.0.2.1.bias", "G.layers.0.0.4.weight", "G.layers.0.0.4.bias", "G.layers.0.0.4.running_mean", "G.layers.0.0.4.running_var", "G.layers.1.0.2.1.weight", "G.layers.1.0.2.1.bias", "G.layers.1.0.4.weight", "G.layers.1.0.4.bias", "G.layers.1.0.4.running_mean", "G.layers.1.0.4.running_var", "G.layers.2.0.2.1.weight", "G.layers.2.0.2.1.bias", "G.layers.2.0.4.weight", "G.layers.2.0.4.bias", "G.layers.2.0.4.running_mean", "G.layers.2.0.4.running_var", "G.layers.3.0.2.1.weight", "G.layers.3.0.2.1.bias", "G.layers.3.0.4.weight", "G.layers.3.0.4.bias", "G.layers.3.0.4.running_mean", "G.layers.3.0.4.running_var", "G.layers.3.2.fn.to_lin_q.weight", "G.layers.3.2.fn.to_lin_kv.net.0.weight", "G.layers.3.2.fn.to_lin_kv.net.1.weight", "G.layers.3.2.fn.to_kv.weight", "G.layers.4.0.2.1.weight", "G.layers.4.0.2.1.bias", "G.layers.4.0.4.weight", "G.layers.4.0.4.bias", "G.layers.4.0.4.running_mean", "G.layers.4.0.4.running_var", "G.layers.5.0.2.1.weight", "G.layers.5.0.2.1.bias", "G.layers.5.0.4.weight", "G.layers.5.0.4.bias", "G.layers.5.0.4.running_mean", "G.layers.5.0.4.running_var", "D.residual_layers.3.1.fn.to_lin_q.weight", "D.residual_layers.3.1.fn.to_lin_kv.net.0.weight", "D.residual_layers.3.1.fn.to_lin_kv.net.1.weight", "D.residual_layers.3.1.fn.to_kv.weight", "D.to_shape_disc_out.1.fn.fn.to_lin_q.weight", "D.to_shape_disc_out.1.fn.fn.to_lin_kv.net.0.weight", "D.to_shape_disc_out.1.fn.fn.to_lin_kv.net.1.weight", "D.to_shape_disc_out.1.fn.fn.to_kv.weight", "D.to_shape_disc_out.3.fn.fn.to_lin_q.weight", "D.to_shape_disc_out.3.fn.fn.to_lin_kv.net.0.weight", "D.to_shape_disc_out.3.fn.fn.to_lin_kv.net.1.weight", "D.to_shape_disc_out.3.fn.fn.to_kv.weight", "GE.layers.0.0.2.1.weight", "GE.layers.0.0.2.1.bias", "GE.layers.0.0.4.weight", "GE.layers.0.0.4.bias", "GE.layers.0.0.4.running_mean", "GE.layers.0.0.4.running_var", "GE.layers.1.0.2.1.weight", "GE.layers.1.0.2.1.bias", "GE.layers.1.0.4.weight", "GE.layers.1.0.4.bias", "GE.layers.1.0.4.running_mean", "GE.layers.1.0.4.running_var", "GE.layers.2.0.2.1.weight", "GE.layers.2.0.2.1.bias", "GE.layers.2.0.4.weight", "GE.layers.2.0.4.bias", "GE.layers.2.0.4.running_mean", "GE.layers.2.0.4.running_var", "GE.layers.3.0.2.1.weight", "GE.layers.3.0.2.1.bias", "GE.layers.3.0.4.weight", "GE.layers.3.0.4.bias", "GE.layers.3.0.4.running_mean", "GE.layers.3.0.4.running_var", "GE.layers.3.2.fn.to_lin_q.weight", "GE.layers.3.2.fn.to_lin_kv.net.0.weight", "GE.layers.3.2.fn.to_lin_kv.net.1.weight", "GE.layers.3.2.fn.to_kv.weight", "GE.layers.4.0.2.1.weight", "GE.layers.4.0.2.1.bias", "GE.layers.4.0.4.weight", "GE.layers.4.0.4.bias", "GE.layers.4.0.4.running_mean", "GE.layers.4.0.4.running_var", "GE.layers.5.0.2.1.weight", "GE.layers.5.0.2.1.bias", "GE.layers.5.0.4.weight", "GE.layers.5.0.4.bias", "GE.layers.5.0.4.running_mean", "GE.layers.5.0.4.running_var", "D_aug.D.residual_layers.3.1.fn.to_lin_q.weight", "D_aug.D.residual_layers.3.1.fn.to_lin_kv.net.0.weight", "D_aug.D.residual_layers.3.1.fn.to_lin_kv.net.1.weight", "D_aug.D.residual_layers.3.1.fn.to_kv.weight", "D_aug.D.to_shape_disc_out.1.fn.fn.to_lin_q.weight", "D_aug.D.to_shape_disc_out.1.fn.fn.to_lin_kv.net.0.weight", "D_aug.D.to_shape_disc_out.1.fn.fn.to_lin_kv.net.1.weight", "D_aug.D.to_shape_disc_out.1.fn.fn.to_kv.weight", "D_aug.D.to_shape_disc_out.3.fn.fn.to_lin_q.weight", "D_aug.D.to_shape_disc_out.3.fn.fn.to_lin_kv.net.0.weight", "D_aug.D.to_shape_disc_out.3.fn.fn.to_lin_kv.net.1.weight", "D_aug.D.to_shape_disc_out.3.fn.fn.to_kv.weight". Unexpected key(s) in state_dict: "G.layers.0.0.2.weight", "G.layers.0.0.2.bias", "G.layers.0.0.3.bias", "G.layers.0.0.3.running_mean", "G.layers.0.0.3.running_var", "G.layers.0.0.3.num_batches_tracked", "G.layers.1.0.2.weight", "G.layers.1.0.2.bias", "G.layers.1.0.3.bias", "G.layers.1.0.3.running_mean", "G.layers.1.0.3.running_var", "G.layers.1.0.3.num_batches_tracked", "G.layers.2.0.2.weight", "G.layers.2.0.2.bias", "G.layers.2.0.3.bias", "G.layers.2.0.3.running_mean", "G.layers.2.0.3.running_var", "G.layers.2.0.3.num_batches_tracked", "G.layers.3.0.2.weight", "G.layers.3.0.2.bias", "G.layers.3.0.3.bias", "G.layers.3.0.3.running_mean", "G.layers.3.0.3.running_var", "G.layers.3.0.3.num_batches_tracked", "G.layers.3.2.fn.to_kv.net.0.weight", "G.layers.3.2.fn.to_kv.net.1.weight", "G.layers.4.0.2.weight", "G.layers.4.0.2.bias", "G.layers.4.0.3.bias", "G.layers.4.0.3.running_mean", "G.layers.4.0.3.running_var", "G.layers.4.0.3.num_batches_tracked", "G.layers.5.0.2.weight", "G.layers.5.0.2.bias", "G.layers.5.0.3.bias", "G.layers.5.0.3.running_mean", "G.layers.5.0.3.running_var", "G.layers.5.0.3.num_batches_tracked", "D.residual_layers.3.1.fn.to_kv.net.0.weight", "D.residual_layers.3.1.fn.to_kv.net.1.weight", "D.to_shape_disc_out.1.fn.fn.to_kv.net.0.weight", "D.to_shape_disc_out.1.fn.fn.to_kv.net.1.weight", "D.to_shape_disc_out.3.fn.fn.to_kv.net.0.weight", "D.to_shape_disc_out.3.fn.fn.to_kv.net.1.weight", "GE.layers.0.0.2.weight", "GE.layers.0.0.2.bias", "GE.layers.0.0.3.bias", "GE.layers.0.0.3.running_mean", "GE.layers.0.0.3.running_var", "GE.layers.0.0.3.num_batches_tracked", "GE.layers.1.0.2.weight", "GE.layers.1.0.2.bias", "GE.layers.1.0.3.bias", "GE.layers.1.0.3.running_mean", "GE.layers.1.0.3.running_var", "GE.layers.1.0.3.num_batches_tracked", "GE.layers.2.0.2.weight", "GE.layers.2.0.2.bias", "GE.layers.2.0.3.bias", "GE.layers.2.0.3.running_mean", "GE.layers.2.0.3.running_var", "GE.layers.2.0.3.num_batches_tracked", "GE.layers.3.0.2.weight", "GE.layers.3.0.2.bias", "GE.layers.3.0.3.bias", "GE.layers.3.0.3.running_mean", "GE.layers.3.0.3.running_var", "GE.layers.3.0.3.num_batches_tracked", "GE.layers.3.2.fn.to_kv.net.0.weight", "GE.layers.3.2.fn.to_kv.net.1.weight", "GE.layers.4.0.2.weight", "GE.layers.4.0.2.bias", "GE.layers.4.0.3.bias", "GE.layers.4.0.3.running_mean", "GE.layers.4.0.3.running_var", "GE.layers.4.0.3.num_batches_tracked", "GE.layers.5.0.2.weight", "GE.layers.5.0.2.bias", "GE.layers.5.0.3.bias", "GE.layers.5.0.3.running_mean", "GE.layers.5.0.3.running_var", "GE.layers.5.0.3.num_batches_tracked", "D_aug.D.residual_layers.3.1.fn.to_kv.net.0.weight", "D_aug.D.residual_layers.3.1.fn.to_kv.net.1.weight", "D_aug.D.to_shape_disc_out.1.fn.fn.to_kv.net.0.weight", "D_aug.D.to_shape_disc_out.1.fn.fn.to_kv.net.1.weight", "D_aug.D.to_shape_disc_out.3.fn.fn.to_kv.net.0.weight", "D_aug.D.to_shape_disc_out.3.fn.fn.to_kv.net.1.weight". size mismatch for G.layers.0.0.3.weight: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for G.layers.1.0.3.weight: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for G.layers.2.0.3.weight: copying a param with shape torch.Size([512]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for G.layers.3.0.3.weight: copying a param with shape torch.Size([256]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for G.layers.3.2.fn.to_out.weight: copying a param with shape torch.Size([256, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([256, 1024, 1, 1]). size mismatch for G.layers.4.0.3.weight: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for G.layers.5.0.3.weight: copying a param with shape torch.Size([64]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for D.residual_layers.3.1.fn.to_out.weight: copying a param with shape torch.Size([128, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([128, 1024, 1, 1]). size mismatch for D.to_shape_disc_out.1.fn.fn.to_out.weight: copying a param with shape torch.Size([64, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([64, 1024, 1, 1]). size mismatch for D.to_shape_disc_out.3.fn.fn.to_out.weight: copying a param with shape torch.Size([32, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([32, 1024, 1, 1]). size mismatch for GE.layers.0.0.3.weight: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for GE.layers.1.0.3.weight: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for GE.layers.2.0.3.weight: copying a param with shape torch.Size([512]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for GE.layers.3.0.3.weight: copying a param with shape torch.Size([256]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for GE.layers.3.2.fn.to_out.weight: copying a param with shape torch.Size([256, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([256, 1024, 1, 1]). size mismatch for GE.layers.4.0.3.weight: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for GE.layers.5.0.3.weight: copying a param with shape torch.Size([64]) from checkpoint, the shape in current model is torch.Size([1]). size mismatch for D_aug.D.residual_layers.3.1.fn.to_out.weight: copying a param with shape torch.Size([128, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([128, 1024, 1, 1]). size mismatch for D_aug.D.to_shape_disc_out.1.fn.fn.to_out.weight: copying a param with shape torch.Size([64, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([64, 1024, 1, 1]). size mismatch for D_aug.D.to_shape_disc_out.3.fn.fn.to_out.weight: copying a param with shape torch.Size([32, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([32, 1024, 1, 1]).

Hi,

thank you for this repo, I've been playing with it a bit and it seems very good! I am trying to generate greyscale images, so I modified the channel accordingly

init_channel = 4 if transparent else 1

unfortunately, this seemed to have no effect as the images generated are still RGB (even though they converge towards greyscale with time), even weirder IMO is that I can modify the number of channels for the generator and keep the original 3 for the discriminator without any issue.

I have also changed this part to no effect

convert_image_fn = partial(convert_image_to, 'RGBA' if transparent else 'L') num_channels = 1 if not transparent else 4

Am I missing something here?

I've been able to train models before but after changing my dataset I'm getting the error.

My trace: File "/usr/local/lib/python3.6/dist-packages/lightweight_gan/lightweight_gan.py", line 1356, in load name = checkpoints[-1] TypeError: 'NoneType' object is not subscriptable

Hello,

First of all, thank you for sharing your code and insights with the rest of us!

As for your code, I plan to run it for 12 hours on Google Colab, similarly to the set-up for what is shown in the README.

My datasets consists of images of 256x256 resolution, and I have started training with the following command-line:

!lightweight_gan \
 --data {image_dir} \
 --disc-output-size 5 \
 --aug-prob 0.25 \
 --aug-types [translation,cutout,color] \
 --amp \

I have noticed that the expected training time is 112.5 hours with 150k iterations (the default setting), which is consistent with the average time of 2.7 seconds per iteration shown in the log. However, it is ~ 9 times more than what is shown in the README. So I wonder if I am doing something wrong, and I see 2 solutions.

First, I could decrease the number of iterations so that it takes 12 hours, by choosing 16k iterations instead of 150k with:

 --num-train-steps 16000 \

Is it what you have done for the results shown in the README?

Second, I have noticed that I am only using 3.8 GB of GPU memory, so I could increase the batch size, as you mentioned in https://github.com/lucidrains/lightweight-gan/issues/13#issuecomment-732486110. Edit: However, the training time increases with a larger batch size. For instance, I am using 7.2 GB of GPU memory, and it takes 8.2 seconds per iteration, with the following:

 --batch-size 32 \
 --gradient-accumulate-every 4 \

Added Experimentation Tracking using Aim.

Now you can:

Track all the model hyperparameters and architectural choices. Track all types of losses. Filter all the experiments with respect to hyperparameters or the architecture Group and aggregate w.r.t. all the trackables to dive into granular experimentation assessment. Track the generated images to track how the model improves.

I'm trying to run the generator after training, to generate fake samples using the following command

lightweight_gan --generate --load-from 299

I get this following error:

Traceback (most recent call last):
  File "C:\anaconda3\lib\runpy.py", line 197, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "C:\anaconda3\lib\runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "C:\anaconda3\Scripts\lightweight_gan.exe\__main__.py", line 7, in <module>
  File "C:\anaconda3\lib\site-packages\lightweight_gan\cli.py", line 195, in main
    fire.Fire(train_from_folder)
  File "C:\anaconda3\lib\site-packages\fire\core.py", line 141, in Fire
    component_trace = _Fire(component, args, parsed_flag_args, context, name)
  File "C:\anaconda3\lib\site-packages\fire\core.py", line 466, in _Fire
    component, remaining_args = _CallAndUpdateTrace(
  File "C:\anaconda3\lib\site-packages\fire\core.py", line 681, in _CallAndUpdateTrace
    component = fn(*varargs, **kwargs)
  File "C:\anaconda3\lib\site-packages\lightweight_gan\cli.py", line 158, in train_from_folder
    model = Trainer(**model_args)
  File "C:\anaconda3\lib\site-packages\lightweight_gan\lightweight_gan.py", line 1057, in __init__
    self.run = self.aim.Run(run_hash=aim_run_hash, repo=aim_repo)
AttributeError: 'Trainer' object has no attribute 'aim'

and when I try to run pip install aim, I get a dependency error with aimrocks

  ERROR: Command errored out with exit status 1:
   command: 'C:\Anaconda3\envs\aerialweb\python.exe' 'C:\Anaconda3\envs\aerialweb\lib\site-packages\pip' install --ignore-installed --no-user --prefix 'C:\Users\ahmed\AppData\Local\Temp\pip-build-env-b2ysw94t\overlay' --no-warn-script-location --no-binary :none: --only-binary :none: -i https://pypi.org/simple -- setuptools 'cython >= 3.0.0a9' 'aimrocks == 0.2.1'
       cwd: None
  Complete output (12 lines):
  WARNING: Ignoring invalid distribution -pencv-python (c:\anaconda3\envs\aerialweb\lib\site-packages)
  WARNING: Ignoring invalid distribution -cipy (c:\anaconda3\envs\aerialweb\lib\site-packages)
  Collecting setuptools
    Using cached setuptools-59.6.0-py3-none-any.whl (952 kB)
  Collecting cython>=3.0.0a9
    Using cached Cython-3.0.0a10-py2.py3-none-any.whl (1.1 MB)
  ERROR: Could not find a version that satisfies the requirement aimrocks==0.2.1 (from versions: 0.1.3a14, 0.2.0.dev1, 0.2.0)
  ERROR: No matching distribution found for aimrocks==0.2.1
  WARNING: Ignoring invalid distribution -pencv-python (c:\anaconda3\envs\aerialweb\lib\site-packages)
  WARNING: Ignoring invalid distribution -cipy (c:\anaconda3\envs\aerialweb\lib\site-packages)
  WARNING: Ignoring invalid distribution -pencv-python (c:\anaconda3\envs\aerialweb\lib\site-packages)
  WARNING: Ignoring invalid distribution -cipy (c:\anaconda3\envs\aerialweb\lib\site-packages)

What is aimrocks and what does it actually do? I am unable to find a matching distribution or even a wheel file to install it manually. Please help

Hello, I hope it's not too much of a noob question, I don't have any background in coding.

After creating the env and installing Pytorch I ran "python setup.py install" and then I ran "python lightweight_gan --data /source --image-size 512" (I filled a "source" folder with pictures of fishes) but I get the error "can't find 'main' module". More exactly, C:\Programmes perso\Logiciels\Anaconda\envs\lightweightgan\python.exe: can't find 'main' module in 'C:\Programmes perso\Logiciels\LightweightGan\lightweight_gan' I tried to copy and rename some of the other modules (init, lightweight_gan...), the code seems to start to run but stops before doing anything. So I guess some file must be missing, or did I do something wrong ?

Thanks a lot for the repo and have a nice day

Hello! Training on Google Colab with

!lightweight_gan --data my/images/ --name my-name --image-size 256 --transparent --dual-contrast-loss --num-train-steps 250000

I'm at 250k iterations over the course of 5 days at 2s/it, and have gotten strange results with boxes.

I've circled some examples of this below.

My training data is 22k images of 256x256 .pngs that do not contain large hard edges or boxes like this. They're video game sprites with hard edges being limited to at most 10x10px

Are there any suggestions I can do with arguments in order to decrease the chance of the models learning that transparent boxes are good? Would converting to a white background help?

Thank you!

If I bring down the number of training steps from 150 000 to 30 000, will the trained model be overall bad? Does it really need the 100 000 or 150 000 training steps?

A rather frustrating issue:

calling it with a trailing \ like lightweight_gan --data full_cleaned256/ --name c256 --models_dir models --results_dir results \

sets the variable new to the truthy value '\' and deletes all progress.

This might well be an issue with Fire but might be mitigated or fixed here too, I am unsure about that.

Thanks. Jonas

Is there any way to reverse engineer the generated images into the latent space?

I am trying to embed fresh RGB as well as ones generated by the Generator into the latent space so I can find its nearest neighbour, pretty much like AI image editing tools.

I plan to convert my RGB image into tensor embeddings based on my trained model and tweak the feature vectors.

How can I achieve this with lightweight-gan?

I am trying to train with a custom image dataset for about 600,000 epochs. At about halfway, my D_loss converges to 0 while my G_loss stays put at 2.5

My evaluation outputs are slowly starting to fade out to either black or white.

Is there any thing that I could to tweak my model? Either by increasing the threshold for the Discriminator or by training the Generator only?

Hi,

Thanks for this amazing implementation! I have a question concerning the loss implementation, as it seems to differ from the original equations. The screenshot below shows the GAN loss as presented in the paper :

• in red, the discriminator loss (D loss) on the true labels,
• in green the D loss on labels for fake generated images,
• and in blue, the generator loss (G loss) on labels for fake images.

This makes sense to me. Since it is assumed that D outputs values between 0 and 1 (0 = fake, 1 = real) :

• in red, we want D to output 1 for true images → let's assume D indeed outputs 1 for true images : -min(0, -1 + D(x)) = 0, which is indeed the minimum achievable,
• in green, we want D to output 0 (from the discriminator perspective) for fake images → let's assume D indeed outputs 0 for fake images : -min(0, -1 - D(x^)) = 1, which is the minimum achievable if D outputs values only between 0 and 1,
• in blue, we want D to output 1 (from the generator perspective) for fake images : the equation follows directly.

Now, the way the authors implement this in the code provided in the supplementary materials of the paper is as follows (the colors match the ones in the above picture)

Except for the strange involved randomness (already explained in https://github.com/lucidrains/lightweight-gan/issues/11), their implementation is a one to one match with the paper equations.

The way it is implemented in this repo however is quite different, and I do not understand why..

Let's start with the discriminator loss :

• in red, you want D to output small values (negative if allowed), to set this term as small as possible (0 if D can output negative values)
• in green, you want D to output values as large as possible (larger or equal to 1) to cancel this term out as well

For the generator loss :

• in blue, you want the opposite of green, that is for D to output values as small as possible

This implementation seems to be meaningful, and yields coherent results (as proven in examples). It also seems to me that D is not limited to output values between 0 and 1, but any real value (I might be wrong). I am just wondering why this choice? Could you perhaps elaborate why you decided to implement the loss differently from the original paper?

how to show generator results after every epoch during training ?

this is my current configuration

 lightweight_gan \
  --data "/content/dataset/Dataset/" \
  --num-train-steps 100000 \
  --image-size 128 \
  --name GAN2DBlood5k \
  --batch-size 32 \
  --gradient-accumulate-every 5 \
  --disc-output-size 1 \
  --dual-contrast-loss \
  --attn-res-layers [] \
  --calculate_fid_every 1000\
  --greyscale \
  --amp

using --show-progress only works after training. Also it seems that there is no longer checkpoints per epoch