从代码上来看，grid表示归一化后的位置信息，范围是0-1 那么这里除以2 再加上0.25的操作怎么理解？ 同时回归的宽高也除以2 https://github.com/Huage001/PaintTransformer/blob/7960b8eef2f6af7171af01933a3d8dbc73190a8d/inference/inference.py#L476 https://github.com/Huage001/PaintTransformer/blob/7960b8eef2f6af7171af01933a3d8dbc73190a8d/inference/inference.py#L477

Hi, I wanted to try to play with so I cloned it locally and run inference but seems I can't fit GPU memory even it shall be fine as the model askes for 3.5GB and I have RTX 2070 with 8GB memory which is completely free...

python inference.py
Traceback (most recent call last):
  File "inference.py", line 490, in <module>
    main(input_path='input/chicago.jpg',
  File "inference.py", line 445, in main
    final_result = param2img_parallel(param, decision, meta_brushes, final_result)
  File "inference.py", line 263, in param2img_parallel
    valid_foregrounds, valid_alphas = param2stroke(param[decision, :], patch_size_y, patch_size_x, meta_brushes)
  File "inference.py", line 74, in param2stroke
    foreground = morphology.dilation(foreground)
  File "/home/jirka/Workingspace/PaintTransformer/inference/morphology.py", line 49, in dilation
    channel = nn.functional.unfold(x_pad, 2 * m + 1, padding=0, stride=1).view(b, c, -1, h, w)
  File "/home/jirka/.local/lib/python3.8/site-packages/torch/nn/functional.py", line 4472, in unfold
    return torch._C._nn.im2col(input, _pair(kernel_size), _pair(dilation), _pair(padding), _pair(stride))
RuntimeError: CUDA out of memory. Tried to allocate 3.38 GiB (GPU 0; 7.80 GiB total capacity; 3.39 GiB already allocated; 3.25 GiB free; 3.41 GiB reserved in total by PyTorch)

well, it seems like it needs to allocate twice the amount, so maybe allow also CPU inference?

Thank you for your wonderful work.

I didn't have access to Google Drive. Will the Pretrained model be published later? Or do you need a request for access?

Firstly, I am very impressed with your project. Great work!

Current behavior

Currently, the user has to open the inference.py file and manually change the location of the input file and other necessary parameters.

Expected behavior

The user can provide the script arguments from the console and does not need to edit the inference.py file.

For example:

python inference.py serial=True

I am happy to contribute and add this change if you see value in it. :)

Firstly, I am very impressed with your project. Great work!

Current behavior

Currently, a potential user of the repository has to install each necessary library separately

Expected behavior

The required libraries would be listed in the requirements.txt file and then the necessary python environment could be installed with a single command.

pip install -r requirements.txt

I am happy to contribute and add this change if you see value in it. :)

hi, great work! I am just wondering how is stylized painting generated? In the paper it was simply put "utilize existing style transfer methods", does it mean you first stylize the original picture to "stylized picture" then convert to painting with your work? or there is something to do with your stylized renderer?

If it is the former one, I suggest you put the stylized picture alongside with the final result to avoid any confusion. If it is the latter one, can you share some more details?

Thanks,

At line 473 of inference.py, it writes

decision = stroke_decision.view(1, h, w, stroke_num).contiguous().bool()

However, this float to bool convert seems to behave wrongly. It almost always yields all true decisions. (I didn't find a document that specifies the converting rule in PyTorch. However, in other languages, float with non-zero values are treated as true.) Consequently, this causes so-called lattice artifacts in the final result.

The following code fixes this issue. :hugs:

decision = stroke_decision.view(1, h, w, stroke_num).contiguous() > 0