search

SwinTransformer / Swin-Transformer-Object-Detection

This is an official implementation for "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" on Object Detection and Instance Segmentation.
https://arxiv.org/abs/2103.14030
Apache License 2.0
1.76k stars 374 forks source link

about swin pre-trained model #47

Open NEUdeep opened 3 years ago

NEUdeep commented 3 years ago

2021-06-01 01:32:06,714 - mmdet - INFO - load model from: checkpoints/cascade_mask_rcnn_swin_base_patch4_window7.pth Traceback (most recent call last): File "tools/train.py", line 187, in main() File "tools/train.py", line 161, in main test_cfg=cfg.get('test_cfg')) File "/workspace/mnt/storage/kanghaidong/cloud_project/Swin-Transformer-Object-Detection/mmdet/models/builder.py", line 77, in build_detector return build(cfg, DETECTORS, dict(train_cfg=train_cfg, test_cfg=test_cfg)) File "/workspace/mnt/storage/kanghaidong/cloud_project/Swin-Transformer-Object-Detection/mmdet/models/builder.py", line 34, in build return build_from_cfg(cfg, registry, default_args) File "/usr/local/lib/python3.6/dist-packages/mmcv/utils/registry.py", line 171, in build_from_cfg return obj_cls(**args) File "/workspace/mnt/storage/kanghaidong/cloud_project/Swin-Transformer-Object-Detection/mmdet/models/detectors/cascade_rcnn.py", line 25, in init pretrained=pretrained) File "/workspace/mnt/storage/kanghaidong/cloud_project/Swin-Transformer-Object-Detection/mmdet/models/detectors/two_stage.py", line 48, in init self.init_weights(pretrained=pretrained) File "/workspace/mnt/storage/kanghaidong/cloud_project/Swin-Transformer-Object-Detection/mmdet/models/detectors/two_stage.py", line 68, in init_weights self.backbone.init_weights(pretrained=pretrained) File "/workspace/mnt/storage/kanghaidong/cloud_project/Swin-Transformer-Object-Detection/mmdet/models/backbones/swin_transformer.py", line 595, in init_weights load_checkpoint(self, pretrained, strict=False, logger=logger) File "/workspace/mnt/storage/kanghaidong/cloud_project/Swin-Transformer-Object-Detection/mmcv_custom/checkpoint.py", line 340, in load_checkpoint table_current = model.state_dict()[table_key] KeyError: 'backbone.layers.0.blocks.0.attn.relative_position_bias_table'

pytorch==1.7.0 is any pre-trained weight error? can you provide a swin_transformer pre-trained model's.thanks.

NEUdeep commented 3 years ago

I know.

bfialkoff commented 3 years ago

If you found the solution please share it for the benefit of others.

novice03 commented 3 years ago

Seems like the pretrained weights expects the model to have an explicit backbone layer since all of the weights for the backbone have 'backbone' in them. I simply removed the backbone. from all the keys in the weight dictionary and manually loaded the weights. Maybe another way to solve this is to make a nn.Module object with self.backbone = SwinTransformer(...)

BowieHsu commented 3 years ago

may caused by difference between setting "window_size=7" in Swin classification model and mmdetection config