RuntimeError: shape '[1, 1, 14, 16, 13]' is invalid for input of size 3120

[EricLe-dev]

Dear @Yaoyi-Li,

Thank you so much for this great work. I'm training this on my custom dataset. With all 512x512 images. The model seems to worked well. However, there are images where I received this error:
RuntimeError: shape '[1, 1, 14, 16, 13]' is invalid for input of size 3120

The problem came from this line and this line in the code. Some images worked well, but there are a lot of images that encountered this error.

Can you please give me a small guide on how to fix it? Thank you so much!

[EricLe-dev]

A bit more information, since you have not yet provided the pretrained model "model_best_resnet34_D_nomixup.pth.tar", I used the ResNet34_En_nomixup model from your previous work (GCA-Matting) and it seems to work, at least, the training ran flawlessly. I'm not sure if that is also a source of that problem.

[EricLe-dev]

@Yaoyi-Li I tested your pretrained model HOP-9x9-RI-all-data.pth and this model also encountered exactly the same problem. Can you please give me a guide on how to fix it? Thank you so much!

[Yaoyi-Li]

Did you add the training dataloader by yourself? I think you can check whether the size of input image is divisible by 32 first. And the input image should be larger than (15*32, 15*32). If your image is smaller than (480, 480), you can pad the image or reduce the default embed_radius

[EricLe-dev]

Thank you for your reply. I did add the training data load myself, using the training from your previous work GCA-Matting. All images in my training data have the size of 512x512. I wrote a piece of code to resize the input images to be divisible by 32. However, that problem still persists. I tested with your pretrained model and it happened exactly the same.

Please find attached one of the image and corresponding trimap that I used to test. Please kindly note that all input images were resize to be divisible by 32.

The error was:
RuntimeError: shape '[1, 13, 1, 12, 12]' is invalid for input of size 2160

[EricLe-dev]

@Yaoyi-Li Seems to be a minor bug in my code. Thank you so much. I figured it out.