gfxdisp / mdf Goto Github PK

While I was testing the MDF loss, I found that the loss function uses much more memory than what is proposed in the paper for only one image. What might be the issue?

import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
target = torch.randn(1, 3, 256, 256).to(device).requires_grad_(False)
input = torch.randn(1, 3, 256, 256).to(device).requires_grad_(True)
mdf = MDFLoss("./Ds_Denoising.pth", True)
loss = torch.zeros(1).cuda()
loss_style = torch.zeros(1).cuda()
j = 0
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
a = torch.cuda.memory_reserved(0)
t = mdf(input, target)
t.backward()
b = torch.cuda.memory_reserved(0)
start.record()
for i in range(100):
    t = mdf(input, target)
    t.backward()
end.record()

# Waits for everything to finish running
torch.cuda.synchronize()

print(str(start.elapsed_time(end)/100) + "ms")
print(str((b-a)/1e6) + "MB")

The output of the code above is placed here:

236.59349609375ms
1369.440256MB

Torch version: 2.0.1+cu117
Python version: 3.10.6
OS: Ubuntu 22.04.2 LTS x86_64
Kernel: 5.15.0-76-generic
CPU: AMD Ryzen 3 3100 (8) @ 3.600GHz
GPU: NVIDIA GeForce GTX 1050 Ti

Hi! I liked your approach and wanted to try it for SISR task, however, I am facing issues while loading the weights file when I want to train it on multiple GPUs. I guess the torch.load() directly is causing the problem here

Can you please pass it as a model instance where we can load it as model.load_state_dict(torch.load("...")) ?? That might be helpful if anyone wants to train using multiple GPUs. Also can you please share the opt argparse arguments mentioned in the SinGAN/models.py file??

Please let me know if you have a workaround for this (loading the weights file on multiple GPUs) and it would also be helpful if you could release the training code.

Thanks!

Hey,
I'm trying to load the SISR weights after cloning the project and I got the following error:

ModuleNotFoundError: No module named 'SinGAN'

May the .pth files are courrupted?

Thanks

Hello, I was pretty thrilled to attempt your loss after reading your paper, however, when I tried to adopt the MDF loss by using Denoising.pth it didn't work, it keeps giving me this issue, could you please tell me how to use your loss in another model?

Awesome work and example notebook. Are there any plans to release the training code?

          I have already set the require_grads to False and used torch.no_grad() for the target - for instance y, but the main idea is that it has to compute the grads with respect to each of the layers and inputs by applying the chain rule. So for the task, we have 8 backpropagations at a time - 8 different models. It creates 8 deep copies of the input. I am interested in how you measure the inference and backpropagation time and the memory overhead.

Best regards,
Delyan

Originally posted by @delyan-boychev in #8 (comment)

Hey, thank you for this project, well done!
Do you mind sharing your generator checkpoint (those that were trained based on the MDF loss)?
I want to simply run inference on the JPEG artifacts removal and see the results

Hi, I'm interested in your work "Training a Task-Specific Image Reconstruction Loss". I'm wondering how to train the discriminators used for calculate the MDF loss (the phase 1 mentioned in the paper) ? It seems that the training code is not provided in this repo.

Dear sir, after I read the code, there are a few questions. How to train the generator mentioned in the paper? And how to define the generator mentioned in this paper?