[Feature] Add PoseFormer backbone #1215
Conversation
* update mmvc installation CI and doc * fix lint
* add derepcation message for deploy tools * change import warnings positions * do yapf * do isort
* hrformer * modify cfg * update url and readme for hrformer. * add readme for hrformer papar * modify reaadme * fix publish year Co-authored-by: ly015 <liyining0712@gmail.com>
* clean inference code * LoadImageFromFile supports given img
…b#1161) co-authored-by: ly015 <liyining0712@gmail.com>
…en-mmlab#1214) * switch to open-mmlab/pre-commit-hooks * deprecate .dev_scripts/github/update_copyright.py
* add windows-ci * reduce input size to save memory * fix codecov * reduce input size to save memory * skip voxelpose unittest for Windows * remove win+cuda test
Codecov ReportAttention:
... and 26 files with indirect coverage changes 📢 Thoughts on this report? Let us know! |
| from .base_backbone import BaseBackbone | ||
|
|
||
|
|
||
| class MultiheadAttention(BaseModule): |
There was a problem hiding this comment.
mmcv also has MultiheadAttention module. Can we use that one?
There was a problem hiding this comment.
This MultiheadAttention module is adapted from mmcls. Compared with the MultiheadAttention in mmcv, its counterpart in mmcls is more similar to the implementation in the official PoseFormer repository.
There was a problem hiding this comment.
I will give mmcv Transformer blocks a try. However, if the output and the model weights of the re-implemented PoseFormer based on mmcv Transformer blocks could not fit the official version of PoseFormer, keeping the mmcls Transformer modules might be a good choice.
| num_joints=17, | ||
| in_chans=2, | ||
| embed_dim_ratio=32, | ||
| depth=4, |
There was a problem hiding this comment.
Is the depth for spatial and temporal transformer always the same? If not, do we need to distinguish them?
There was a problem hiding this comment.
In the official implementation of PoseFormer, the spatial and temporal transformer share the same depth.
However, I think it's obviously better to distinguish the two depths for clarity and I will follow this comment.
| if norm_cfg is None: | ||
| norm_cfg = dict(type='LN') | ||
| # Temporal embed_dim is num_joints * spatial embedding dim ratio | ||
| embed_dim = embed_dim_ratio * num_joints |
There was a problem hiding this comment.
Does embed_dim mean the embedding dimension for each frame? How about using embed_dim_per_frame ?
There was a problem hiding this comment.
Now I use spatial_embed_dim and temporal_embed_dim instead of embed_dim_ratio and embed_dim.
| init_cfg=init_cfg) for i in range(depth) | ||
| ]) | ||
|
|
||
| self.blocks = nn.ModuleList([ |
There was a problem hiding this comment.
Does self.blocks mean the temporal transformer blocks? How about using self.temporal_blocks for clarity?
There was a problem hiding this comment.
self.temporal_blocks is a better choice.
| return x | ||
|
|
||
|
|
||
| class TransformerEncoderLayer(BaseModule): |
There was a problem hiding this comment.
mmcv has the similar layer BaseTransformerLayermodule under mmcv/cnn/bricks/transformer.py, which also includes MultiheadAttention. Can we use that one?
There was a problem hiding this comment.
This TransformerEncoderLayer module is adapted from mmcls. Compared with the BaseTransformerLayer in mmcv, its counterpart in mmcls is more similar to the implementation in the official PoseFormer repository.
There was a problem hiding this comment.
I will give mmcv Transformer blocks a try. However, if the output and the model weights of the re-implemented PoseFormer based on mmcv Transformer blocks could not fit the official version of PoseFormer, keeping the mmcls Transformer modules might be a good choice.
Drop path rate / weight_decay
Change samples_per_gpu to 128
|
Sorry, I have resigned from the Shanghai AI Lab for a long time. I'm busy working on other projects so I don't have much time for this. If someone could help me finish this feature, I will be always willing to help. |
Motivation
Modification
Add backbone, head and config of the PoseFormer (ICCV 2021) into the repository.
BC-breaking (Optional)
Use cases (Optional)
Checklist
Before PR:
After PR: