progress bar updating in mmpose/apis/test.py is incorrect for multi-frame detectors such as posewarper

[zwfcrazy]

Describe the bug

Batch size count in single_gpu_test and multi_gpu_test, such as the line here, https://github.com/open-mmlab/mmpose/blob/f8b79fe3d92aa71c29c16835cbe7af11204f412c/mmpose/apis/test.py#L37 is incorrect for multi-frame detectors such as posewarper. This line assumes the input image is a tensor while for posewarper the input is a list of tensors and the length of list is equal to number of frames used for inference not batch size.

Reproduction Use tools/test.py to test a posewarper stage2 model and observe how the progress bar updates.

Environment Irrelevant

Error traceback Not applicable

Bug fix Probably using number of entries in img_metas is a better approach?

[liqikai9]

@zwfcrazy Thanks for your feedback! This is actually a bug. Currently, the input data is a dictionary so we just use the length of the first value in data as the batch size.

Probably using number of entries in img_metas is a better approach?

This may work. Would you like to raise a pull request to fix it by yourself?

[zwfcrazy]

@liqikai9 I am currently using this modification batch_size = len(data['img_metas'].data[0]) for my own test, but I am not sure if this is general enough.

[liqikai9]

I see. It seems like a good choice for this detector and config. We will continue to pay attention to this issue and find a general way to fix it. Thanks again~