A little error report about BaseHead and HeatmapHead class in mmpose 1.0.0 rc

[xinxinxinxu]

I wonder if anyone has encountered a similar problem when using a single-resolution pred map to consulate mseloss , there will report that tensors size mismatch between preds and gt. For instance, the output of the backbone is not a list but a tensor, like (64, 32, 64, 64). When it is processed by the __transform_inputs function(a private function of BaseHead class), it will lose a dimension, like turning to (16, 64, 64).

However, this error does not exist in the previous version. I found this is because there is a conditional statement that was missing in the new version, after adding it, the error will no longer be reported.

Original：

Modified：

Thanks to the developers for their contributions！ Hope this report is helpful for everyone！

[Ben-Louis]

Hi, thanks for using MMPose. In MMPose 1.x, the default output of the backbone is a tuple, with which heads are designed to deal. For example, the raw output of Resnet50 (a tensor) will be packed into a tuple with length 1, and will be unpacked in xxhead._transform_inputs with parameters input_transform='select' and input_index=-1.

But I think your modification makes sense. Because the output of most backbones that are not implemented in MMPose is a tensor. If users want to try some advanced networks in pose estimation with MMPose, they should not be bothered by this adaptation problem. What do you think? @ly015

[ly015]

@xinxinxinxu Thanks for reporting this issue. The modification is indeed necessary for users who would like to implement their own backbone.

Would you be willing to submit a pull request to MMPose dev-1.x branch to fix this issue? A brief guide to contributing codes to MMPose can be found at: https://mmpose.readthedocs.io/en/dev-1.x/notes/contribution_guide.html. Your contribution would be much appreciated.