Bowen Zheng

kd_loss implementation issue

9

Hello, I found that the `knowledge_distillation_kl_div_loss()` in `mmdet/models/losses/kd_loss.py` uses a different implementation compared to the normal KL Div definition, which is equivalent to `F.kl_div(reduction='mean')` instead of `F.kl_div(reduction='batchmean')` as mentioned in...

Question about the hyper-parameters used in other KD methods on different cases

4

First of all, thank you for the excellent work. We are currently attempting to reproduce the performance of various KD methods, including FitNet, RKD, CRD, ReviewKD, and others, as detailed...

Fix coco_stuff_164k dataset issue

2

Fix the wrong class num in coco_stuff_164k: #27.

New interface does not support custom empty pytree class inherited from dict

1

Reproduction code: ```python class PyTreeDict(dict): pass jax.tree_util.register_pytree_node( PyTreeDict, lambda d: (tuple(d.values()), tuple(d.keys())), lambda keys, values: PyTreeDict(dict(zip(keys, values))) ) a={"a": PyTreeDict()} # ValueError: Expected dict, got {}. # a=PyTreeDict() # ValueError:...

## Description This PR fixes - doc typos. - potential issues of type promotion in the state. - test cases in new StdWorkflow API (WIP) ## Checklist - [x] I...

## Description This PR fixes - doc typos. - potential issues of type promotion in the state. - test cases in new StdWorkflow API (WIP) - wrong fitness in brax...