Giters

fab-jul / L3C-PyTorch

PyTorch Implementation of the CVPR'19 Paper "Practical Full Resolution Learned Lossless Image Compression"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

PyTorch 1.2+ Not Supported

fab-jul opened this issue · comments

commented

Will add here what I find. For now:

  • "Comparison operations (lt (<), le (<=), gt (>), ge (>=), eq (==), ne, (!=) ) return dtype has changed from torch.uint8 to torch.bool": This breaks logistic_mixture.py.
  • "sum(Tensor) (python built-in) does not upcast dtype like torch.sum" → Check for this in code.
  • "Tensorboard is no Longer Considered Experimental" → should use built in Tensorboard
commented

AT_CHECK is deprecated, use TORCH_CHECK.

commented

Is there any hope of training on Pytorch1.7? My video card is a series 30

commented

Probably easier to just create an environment with PyTorch 1.1, or is this not possible?

commented

I successfully coded and decoded in pytorch1.7, and the weight trained in pytorch1.7 appeared noise in the images after encoding and decoding.The weight trained on pytorch1.1 is no problem after encoding and decoding with pytorch1.7. Do you have any suggestions for me to find the problem? Thank you very much.

commented

I successfully coded and decoded in pytorch1.7, and the weight trained in pytorch1.7 appeared noise in the images after encoding and decoding.The weight trained on pytorch1.1 is no problem after encoding and decoding with pytorch1.7. Do you have any suggestions for me to find the problem? Thank you very much.

It seems I encountered same problems with you, I have to use pytorch 1.7 with RTX 3090. The decoded PNG files seems to be noisy.
Do you have any solution to this?
000000000802
000000000802

commented

It seems I encountered same problems with you, I have to use pytorch 1.7 with RTX 3090. The decoded PNG files seems to be noisy. Do you have any solution to this?

I tried to find it but couldn't find it. There is probably a problem with the weights saved during training, because it is correct to run forward images on Pytorch1.7 using the weights published by the author.

commented

I successfully coded and decoded in pytorch1.7, and the weight trained in pytorch1.7 appeared noise in the images after encoding and decoding.The weight trained on pytorch1.1 is no problem after encoding and decoding with pytorch1.7. Do you have any suggestions for me to find the problem? Thank you very much.

It seems I encountered same problems with you, I have to use pytorch 1.7 with RTX 3090. The decoded PNG files seems to be noisy. Do you have any solution to this? 000000000802 000000000802

@fab-jul Please Dr. Fabian, Would you please give some comments how to debug this problem(noisy decoded images)?
Should it be a problem with torchac, the encoder/decoder part, or save/restore part?
Thanks a lot.

commented

It seems I encountered same problems with you, I have to use pytorch 1.7 with RTX 3090. The decoded PNG files seems to be noisy. Do you have any solution to this?

I tried to find it but couldn't find it. There is probably a problem with the weights saved during training, because it is correct to run forward images on Pytorch1.7 using the weights published by the author.

I made some experiments between 2 servers, A with pytorch 1.1, RTX 2060, B with pytroch 1.7, RTX 3090.
model A generated from server A
model B generated from server B using same source code with A

model A can be used to encoding/decoding in both server A and B.
model B can not perform encoding/decoding, even in server B, the decoded images are noisy.

I tried to downgrade pytorch to 1.1 in server B, but problem still not solved.
I have no idea how to debug this problem anymore, so there is no hope to run this code in 3090 GPU?