pre-release models and logs for det and seg

Westlake-AI · Feb 9, 2023 · 23d17b6 · 23d17b6
1 parent cb7ae13
commit 23d17b6
Show file tree

Hide file tree

Showing 28 changed files with 285 additions and 76 deletions.
diff --git a/README.md b/README.md
@@ -98,10 +98,12 @@ This project is released under the [Apache 2.0 license](LICENSE).
 
 Our implementation is mainly based on the following codebases. We gratefully thank the authors for their wonderful works.
 
-- [pytorch-image-models](https://github.com/rwightman/pytorch-image-models)
-- [PoolFormer](https://github.com/sail-sg/poolformer)
-- [MMDetection](https://github.com/open-mmlab/mmdetection)
-- [MMSegmentation](https://github.com/open-mmlab/mmsegmentation)
+- [pytorch-image-models](https://github.com/rwightman/pytorch-image-models).
+- [PoolFormer](https://github.com/sail-sg/poolformer): Official PyTorch implementation of MetaFormer.
+- [ConvNeXt](https://github.com/facebookresearch/ConvNeXt): Official PyTorch implementation of ConvNeXt.
+- [MMDetection](https://github.com/open-mmlab/mmdetection): OpenMMLab Detection Toolbox and Benchmark.
+- [MMSegmentation](https://github.com/open-mmlab/mmsegmentation): OpenMMLab Semantic Segmentation Toolbox and Benchmark.
+- [MMPose](https://github.com/open-mmlab/mmpose): OpenMMLab Pose Estimation Toolbox and Benchmark.
 
 ## Citation
 

diff --git a/detection/README.md b/detection/README.md
@@ -5,7 +5,7 @@ For more details, see [Efficient Multi-order Gated Aggregation Network](https://
 
 ## Note
 
-Please note that we just simply follow the hyper-parameters of [PVT](https://github.com/whai362/PVT/tree/v2/detection) which may not be the optimal ones for MogaNet. Feel free to tune the hyper-parameters to get better performance.
+Please note that we simply follow the hyper-parameters of [PVT](https://github.com/whai362/PVT/tree/v2/detection) and [ConvNeXt](https://github.com/facebookresearch/ConvNeXt), which may not be the optimal ones for MogaNet. Feel free to tune the hyper-parameters to get better performance.
 
 ## Environement Setup
 
@@ -47,19 +47,19 @@ Prepare COCO according to the guidelines in [MMDetection](https://github.com/ope
 ## Results and models on COCO
 
 | Method | Backbone | Pretrain | Params | FLOPs | Lr schd | Aug | box mAP | mask mAP | Config | Download |
-|:---:|:---:|:---:|---|---|:---:|:---:|:---:|:---:|:---:|:---:|
-| RetinaNet | MogaNet-XT | ImageNet-1K | 12.1M | 167.2G | 1x | No | 38.9 | - | [config](configs/mask_rcnn_moganet_xtiny_fpn_1x_coco.py) | log / model |
-| RetinaNet | MogaNet-T | ImageNet-1K | 14.4M | 173.4G | 1x | No | 40.9 | - | [config](configs/mask_rcnn_moganet_tiny_fpn_1x_coco.py) | log / model |
-| RetinaNet | MogaNet-S | ImageNet-1K | 35.1M | 253.0G | 1x | No | 45.4 | - | [config](configs/mask_rcnn_moganet_small_fpn_1x_coco.py) | log / model |
-| RetinaNet | MogaNet-B | ImageNet-1K | 53.5M | 354.5G | 1x | No |  | - | [config](configs/mask_rcnn_moganet_base_fpn_1x_coco.py) | log / model |
-| RetinaNet | MogaNet-L | ImageNet-1K | 92.4M | 476.8G | 1x | No |  | - | [config](configs/mask_rcnn_moganet_large_fpn_1x_coco.py) | log / model |
+|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| RetinaNet | MogaNet-XT | ImageNet-1K | 12.1M | 167.2G | 1x | No | 39.7 | - | [config](configs/mask_rcnn_moganet_xtiny_fpn_1x_coco.py) | log / model |
+| RetinaNet | MogaNet-T | ImageNet-1K | 14.4M | 173.4G | 1x | No | 41.4 | - | [config](configs/mask_rcnn_moganet_tiny_fpn_1x_coco.py) | log / model |
+| RetinaNet | MogaNet-S | ImageNet-1K | 35.1M | 253.0G | 1x | No | 45.8 | - | [config](configs/mask_rcnn_moganet_small_fpn_1x_coco.py) | log / model |
+| RetinaNet | MogaNet-B | ImageNet-1K | 53.5M | 354.5G | 1x | No | 47.7 | - | [config](configs/mask_rcnn_moganet_base_fpn_1x_coco.py) | log / model |
+| RetinaNet | MogaNet-L | ImageNet-1K | 92.4M | 476.8G | 1x | No | 48.7 | - | [config](configs/mask_rcnn_moganet_large_fpn_1x_coco.py) | log / model |
 | Mask R-CNN | MogaNet-XT | ImageNet-1K | 22.8M | 185.4G | 1x | No | 40.7 | 37.6 | [config](configs/retinanet_moganet_xtiny_fpn_1x_coco.py) | log / model |
 | Mask R-CNN | MogaNet-T | ImageNet-1K | 25.0M | 191.7G | 1x | No | 42.6 | 39.1 | [config](configs/retinanet_moganet_tiny_fpn_1x_coco.py) | log / model |
-| Mask R-CNN | MogaNet-S | ImageNet-1K | 45.0M | 271.6G | 1x | No | 46.1 | 41.8 | [config](configs/retinanet_moganet_small_fpn_1x_coco.py) | log / model |
-| Mask R-CNN | MogaNet-B | ImageNet-1K | 63.4M | 373.1G | 1x | No | 48.2 | 43.4 | [config](configs/retinanet_moganet_base_fpn_1x_coco.py) | log / model |
-| Mask R-CNN | MogaNet-L | ImageNet-1K | 102.1M | 495.3G | 1x | No |  |  | [config](configs/retinanet_moganet_large_fpn_1x_coco.py) | log / model |
+| Mask R-CNN | MogaNet-S | ImageNet-1K | 45.0M | 271.6G | 1x | No | 46.6 | 42.2 | [config](configs/retinanet_moganet_small_fpn_1x_coco.py) | log / model |
+| Mask R-CNN | MogaNet-B | ImageNet-1K | 63.4M | 373.1G | 1x | No | 49.0 | 43.8 | [config](configs/retinanet_moganet_base_fpn_1x_coco.py) | log / model |
+| Mask R-CNN | MogaNet-L | ImageNet-1K | 102.1M | 495.3G | 1x | No | 49.4 | 44.2 | [config](configs/retinanet_moganet_large_fpn_1x_coco.py) | log / model |
 
-**Notes**: All the models can also be downloaded by [**Baidu Cloud**](https://pan.baidu.com/s/1d5MTTC66gegehmfZvCQRUA?pwd=z8mf) (z8mf). The params (M) and FLOPs (G) are measured by [get_flops](get_flops.sh) with 1280 $\times$ 800 resolutions.
+**Notes**: All the models can also be downloaded by [**Baidu Cloud**](https://pan.baidu.com/s/1d5MTTC66gegehmfZvCQRUA?pwd=z8mf) (z8mf) at `MogaNet/COCO_Detection`. The params (M) and FLOPs (G) are measured by [get_flops](get_flops.sh) with 1280 $\times$ 800 resolutions.
 ```bash
 bash get_flops.sh /path/to/config --shape 1280 800
 ```
@@ -97,6 +97,7 @@ Our implementation is mainly based on the following codebases. We gratefully tha
 
 - [MMDetection](https://github.com/open-mmlab/mmdetection)
 - [PVT detection](https://github.com/whai362/PVT/tree/v2/detection)
+- [ConvNeXt](https://github.com/facebookresearch/ConvNeXt)
 - [PoolFormer](https://github.com/sail-sg/poolformer)
 
 <p align="right">(<a href="#top">back to top</a>)</p>
diff --git a/detection/configs/moganet/mask_rcnn_moganet_base_fpn_1x_coco.py b/detection/configs/moganet/mask_rcnn_moganet_base_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_base_sz224_8xbs128_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_base_sz224_8xbs128_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -23,5 +23,11 @@
         out_channels=256,
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0002, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/mask_rcnn_moganet_large_fpn_1x_coco.py b/detection/configs/moganet/mask_rcnn_moganet_large_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_large_sz224_8xbs64_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_large_sz224_8xbs64_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -23,5 +23,11 @@
         out_channels=256,
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0002, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/mask_rcnn_moganet_small_fpn_1x_coco.py b/detection/configs/moganet/mask_rcnn_moganet_small_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_small_sz224_8xbs128_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_small_sz224_8xbs128_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -23,5 +23,11 @@
         out_channels=256,
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0002, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0002, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/mask_rcnn_moganet_tiny_fpn_1x_coco.py b/detection/configs/moganet/mask_rcnn_moganet_tiny_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_tiny_sz224_8xbs128_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_tiny_sz224_8xbs128_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -23,5 +23,11 @@
         out_channels=256,
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0002, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0002, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/mask_rcnn_moganet_xtiny_fpn_1x_coco.py b/detection/configs/moganet/mask_rcnn_moganet_xtiny_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_xtiny_sz224_8xbs128_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_xtiny_sz224_8xbs128_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -23,5 +23,11 @@
         out_channels=256,
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0002, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0002, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/retinanet_moganet_base_fpn_1x_coco.py b/detection/configs/moganet/retinanet_moganet_base_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_base_sz224_8xbs128_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_base_sz224_8xbs128_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -25,5 +25,11 @@
         add_extra_convs='on_input',
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/retinanet_moganet_large_fpn_1x_coco.py b/detection/configs/moganet/retinanet_moganet_large_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_large_sz224_8xbs64_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_large_sz224_8xbs64_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -25,5 +25,11 @@
         add_extra_convs='on_input',
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/retinanet_moganet_small_fpn_1x_coco.py b/detection/configs/moganet/retinanet_moganet_small_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_small_sz224_8xbs128_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_small_sz224_8xbs128_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -25,5 +25,11 @@
         add_extra_convs='on_input',
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/retinanet_moganet_tiny_fpn_1x_coco.py b/detection/configs/moganet/retinanet_moganet_tiny_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_tiny_sz224_8xbs128_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_tiny_sz224_8xbs128_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -25,5 +25,11 @@
         add_extra_convs='on_input',
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/detection/configs/moganet/retinanet_moganet_xtiny_fpn_1x_coco.py b/detection/configs/moganet/retinanet_moganet_xtiny_fpn_1x_coco.py
@@ -14,7 +14,7 @@
         init_cfg=dict(
             type='Pretrained', 
             checkpoint=\
-                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_xtiny_sz224_8xbs128_ep300.pth.tar', 
+                'https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_xtiny_sz224_8xbs128_ep300.pth.tar',
             ),
         ),
     neck=dict(
@@ -25,5 +25,11 @@
         add_extra_convs='on_input',
         num_outs=5))
 # optimizer
-optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, weight_decay=0.0001)
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, betas=(0.9, 0.999), weight_decay=0.05,
+                 paramwise_cfg=dict(custom_keys={'layer_scale': dict(decay_mult=0.),
+                                                 'scale': dict(decay_mult=0.),
+                                                 'norm': dict(decay_mult=0.)}))
 optimizer_config = dict(grad_clip=None)
+
+checkpoint_config = dict(interval=1, max_keep_ckpts=1)
+evaluation = dict(save_best='auto')
diff --git a/models/moganet.py b/models/moganet.py
@@ -717,6 +717,7 @@ def freeze_patch_emb(self):
     def no_weight_decay(self):
         return dict()
 
+    @torch.jit.ignore
     def get_classifier(self):
         return self.head
 
@@ -750,7 +751,10 @@ def forward_features(self, x):
             return outs
         else:
             # output only the last layer for image classification
-            return x.mean(dim=[2, 3])
+            return x
+
+    def forward_head(self, x):
+        return self.head(x.mean(dim=[2, 3]))
 
     def forward(self, x):
         x = self.forward_features(x)
@@ -759,8 +763,7 @@ def forward(self, x):
             return x
         else:
             # for image classification
-            x = self.head(x)
-            return x
+            return self.forward_head(x)
 
 
 def _cfg(url='', **kwargs):

diff --git a/pose_estimation/README.md b/pose_estimation/README.md
@@ -5,7 +5,7 @@ For more details, see [Efficient Multi-order Gated Aggregation Network](https://
 
 ## Note
 
-Please note that we just simply follow the hyper-parameters of [PVT](https://github.com/whai362/PVT/tree/v2/detection) and [Swin](https://github.com/microsoft/Swin-Transformer) which may not be the optimal ones for MogaNet. Feel free to tune the hyper-parameters to get better performance.
+Please note that we simply follow the hyper-parameters of [PVT](https://github.com/whai362/PVT/tree/v2/detection) and [Swin](https://github.com/microsoft/Swin-Transformer) which may not be the optimal ones for MogaNet. Feel free to tune the hyper-parameters to get better performance.
 
 ## Environement Setup
 
@@ -26,7 +26,7 @@ Prepare COCO according to the guidelines in [MMPose](https://github.com/open-mml
 
 ## Results and models on COCO
 
-**Notes**: All the models can also be downloaded by [**Baidu Cloud**](https://pan.baidu.com/s/1d5MTTC66gegehmfZvCQRUA?pwd=z8mf) (z8mf). The params (M) and FLOPs (G) are measured by [get_flops](get_flops.sh) with 256 $\times$ 192 or 384 $\times$ 288 resolutions.
+**Notes**: All the models can also be downloaded by [**Baidu Cloud**](https://pan.baidu.com/s/1d5MTTC66gegehmfZvCQRUA?pwd=z8mf) (z8mf) at `MogaNet/COCO_Pose`. The params (M) and FLOPs (G) are measured by [get_flops](get_flops.sh) with 256 $\times$ 192 or 384 $\times$ 288 resolutions.
 ```bash
 bash get_flops.sh /path/to/config --shape 256 192
 ```

diff --git a/...timation/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/moganet_s_coco_256x192.py b/...timation/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/moganet_s_coco_256x192.py
@@ -38,7 +38,7 @@
         arch="small",  # modify 'arch' for various architectures
         init_value=1e-5,
         frozen_stages=1,
-        drop_path_rate=0.1,
+        drop_path_rate=0.2,
         stem_norm_cfg=norm_cfg,
         conv_norm_cfg=norm_cfg,
         out_indices=(0, 1, 2, 3),