Inference error using pretrained TinaFace model on my png file

[junwei-h]

Hello, On Windows WSL, compiled and installed vedadet successfully. When using my own image 128x128 png file to test TinaFace, I got the following error:

Expected number of channels in input to be divisible by num_groups, but got input of shape [64, 560, 560] and num_groups=32

Python 3.8
torch                     1.13.1

No GPU, I ran

python infer.py configs/infer/tinaface/tinaface_r50_fpn_gn_dcn.py test.png

The pre-trained model is tinaface_r50_fpn_gn_dcn.pth and the test.png has 128 x 128 px.

Here is how far I traced the error & the image dimension:

• At https://github.com/Media-Smart/vedadet/blob/54731b9abdc3ca5dcbd9c95fb8abc8f76bad6227/tools/infer.py#L88, I have

  data['img'].shape 
  torch.Size([1, 3, 1120, 1120])

• At https://github.com/Media-Smart/vedadet/blob/54731b9abdc3ca5dcbd9c95fb8abc8f76bad6227/vedadet/engines/infer_engine.py#L57 I have

  img.shape 
  torch.Size([3, 1120, 1120])

• At https://github.com/Media-Smart/vedadet/blob/54731b9abdc3ca5dcbd9c95fb8abc8f76bad6227/vedadet/models/backbones/resnet.py#L624 I have

  x.shape
  torch.Size([64, 560, 560])

• At https://github.com/pytorch/pytorch/blob/5913437a40a6e45ab7e164afb7c6ec930dd40b2f/torch/nn/modules/normalization.py#L273 It triggered the error as I have

  input.shape 
  torch.Size([64, 560, 560])

  Do I need to change 'num_groups=32' in tinaface_r50_fpn_gn_dcn.py?

Any suggestions? Thank you.