Dear author(sniklaus)

i run the code run.py and in the estimate->network->init, it has a state_dict_url named 'http://content.sniklaus.com/sepconv/network'

but i cannot enter this website. Firstly i consider my internet refuse to this web, but my friends also cannot enter this web. Could u plz tell me how to enter this web, or could u plz give me another website about the pretrained model. Thanks a lot

Dear author, I want to use your model to deal with 1624 × 1224 resolution video, but unfortunately, your model seems to limit the maximum resolution to 1280 × 720。 How can you change your code to deal with higher resolution video? I'm a little worried. If you see the message, please contact my email: [email protected]

Hi, I was trying to retrain your model as would like to conduct some experiments on this, however, I've got a few questions about the training process of SepConv++,

For training:

• You train using batch size 16, Adamax optimizer with lr=0.001 which halves at epoch 60 and 80 and train on vimeo without any validation set
• In the paper you said you use relu1_2 of VGG, but you didn't specify which one, is it VGG16 or 19?
• I used VGG19 and defined it as the following:

vgg19 = torchvision.models.vgg19(pretrained=True)
self.vgg = torch.nn.Sequential(*list(vgg19.children())[0][:4])
for param in self.vgg.parameters():
param.requires_grad = False

and to get contextual output:

tenOut2 = sepconv.sepconv_func.apply(self.vgg(tenOne), tenVerone, tenHorone) + sepconv.sepconv_func.apply(self.vgg(tenTwo), tenVertwo, tenHortwo) This gives a 64 channel output. Is this correct? When using the VGG, you can’t add the additional channel to the input to make it 4 channels, so I proceed to apply context on the 3 channel input images

-Do you normalize this output as well? If you do, you get 63 channels unless you remove [:, -1:, :, :]. So here's what I did to normalize without losing that channel:

tenNormalize = tenOut2
tenNormalize[tenNormalize.abs() < 0.01] = 1.0
tenOut2 = tenOut2/ tenNormalize

• When you normalize the output, is this done only during testing or also during training?

• The contextual loss is L1(output,gt) + 0.1*L1(contextual_output, contextual_gt)?

Your help would be appreciated on this as the model isn't converging which leads me to believe I've made some sort of mistake somewhere

Hi @sniklaus, a great work of adaptive conv. However, I have met some issues when copying your code directly to other framework. Could you help me solve them? Thx ahead.

• RuntimeError: CUDA error: an illegal memory access was encountered When I try to clip the result from sepconv as yours here. There is an Error:RuntimeError: CUDA error: an illegal memory access was encountered. I am not familiar with CUDA coding. Could you offer your help about this?
• RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED when I try to backward the gradient, I encounter another Error like this: RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED. I wonder if I missed something important when training sepconv. Thanks for your time!

With the following configuration:

Windows 10
Python 3.8
cupy-cuda11x                 11.0.0
torch                        1.12.1+cu116
NVidia GPU Computing Toolkit 11.7

I get the warning C:\Program Files\Python\Python38\lib\site-packages\cupy\cuda\compiler.py:460: UserWarning: cupy.cuda.compile_with_cache has been deprecated in CuPy v10, and will be removed in the future. Use cupy.RawModule or cupy.RawKernel instead. when I run python run.py --model paper --one ./images/one.png --two ./images/two.png --out ./out.png Any thoughts on how to fix this?

Hi,

When I tried to run this line: self.load_state_dict(torch.hub.load_state_dict_from_url(url='http://content.sniklaus.com/resepconv/network-' + arguments_strModel + '.pytorch', file_name='resepconv-' + arguments_strModel))

I got this error: urllib.error.HTTPError: HTTP Error 404: Not Found

Could you check if this link still works properly?

Thanks!