search

JDAI-CV / fast-reid

SOTA Re-identification Methods and Toolbox
Apache License 2.0
3.39k stars 830 forks source link

Missing Softmax for accuracy logging #694

Closed wobbba closed 1 year ago

wobbba commented 1 year ago

In the head embedding_head.py we output cls_outputs and pred_class_logits. In the meta_arch baseline.py the pred_class_logits are used for accuracy logging while cls_outputs are used as input for CE loss. When using e.g. the provided bagtricks config, we use any_softmax.Linear as cls_layer. This class seems to be unfinished or just wrong. As long as scale=1 (default) it is just the identity. Instead the linear layer is hardcoded into embedding_head using a weight Tensor. What is this mess?

Anyway, who am I to judge this spaghetti code... The main problem is that as a result of this mess, pred_class_logits and cls_outputs are both just the plain logits which is fine for cls_outputs as this is used for CE, where softmax is applied but not for pred_class_logits because in the logging function there is no softmax.

A greater reworking of embedding_head seems in order but for a hot-fix, one can simply manually add softmax in embedding_head pred_class_logits output. (Have not tested yet but will update once I have, if I don't forget)

Or is there something I am missing?

github-actions[bot] commented 1 year ago

This issue is stale because it has been open for 30 days with no activity.

github-actions[bot] commented 1 year ago

This issue was closed because it has been inactive for 14 days since being marked as stale.