[Feature] Support Resnet strikes back

configs/pspnet/README.md

@@ -32,6 +32,15 @@ Scene parsing is challenging for unrestricted open vocabulary and diverse scenes
 }
 ```
 
+```bibtex
+@article{wightman2021resnet,
+  title={Resnet strikes back: An improved training procedure in timm},
+  author={Wightman, Ross and Touvron, Hugo and J{\'e}gou, Herv{\'e}},
+  journal={arXiv preprint arXiv:2110.00476},
+  year={2021}
+}
+```
+
 ## Results and models
 
 ### Cityscapes
@@ -44,6 +53,7 @@ Scene parsing is challenging for unrestricted open vocabulary and diverse scenes
 | PSPNet | R-101-D8  | 769x769   |   40000 | 10.9     | 1.15           | 79.08 |         80.28 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r101-d8_769x769_40k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101-d8_769x769_40k_cityscapes/pspnet_r101-d8_769x769_40k_cityscapes_20200606_112753-61c6f5be.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101-d8_769x769_40k_cityscapes/pspnet_r101-d8_769x769_40k_cityscapes_20200606_112753.log.json)         |
 | PSPNet | R-18-D8   | 512x1024  |   80000 | 1.7      | 15.71          | 74.87 |         76.04 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r18-d8_512x1024_80k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r18-d8_512x1024_80k_cityscapes/pspnet_r18-d8_512x1024_80k_cityscapes_20201225_021458-09ffa746.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r18-d8_512x1024_80k_cityscapes/pspnet_r18-d8_512x1024_80k_cityscapes-20201225_021458.log.json)         |
 | PSPNet | R-50-D8   | 512x1024  |   80000 | -        | -              | 78.55 |         79.79 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes/pspnet_r50-d8_512x1024_80k_cityscapes_20200606_112131-2376f12b.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes/pspnet_r50-d8_512x1024_80k_cityscapes_20200606_112131.log.json)         |
+| PSPNet | R-50-D8 rsb  | 512x1024 | 80000 | 6.2   | 3.82          | 78.47 |         79.45 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r50-d8_rsb-pretrain_512x1024_adamw_80k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes/pspnet_r50-d8_rsb-pretrain_512x1024_adamw_80k_cityscapes_20220315_123238-588c30be.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes/pspnet_r50-d8_rsb-pretrain_512x1024_adamw_80k_cityscapes_20220315_123238.log.json)         |
 | PSPNet | R-101-D8  | 512x1024  |   80000 | -        | -              | 79.76 |         81.01 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r101-d8_512x1024_80k_cityscapes.py)  | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101-d8_512x1024_80k_cityscapes/pspnet_r101-d8_512x1024_80k_cityscapes_20200606_112211-e1e1100f.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101-d8_512x1024_80k_cityscapes/pspnet_r101-d8_512x1024_80k_cityscapes_20200606_112211.log.json)     |
 | PSPNet (FP16) | R-101-D8 | 512x1024  |   80000 | 5.34     | 8.77           | 79.46 |             - | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r101-d8_fp16_512x1024_80k_cityscapes.py)        | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101-d8_fp16_512x1024_80k_cityscapes/pspnet_r101-d8_fp16_512x1024_80k_cityscapes_20200717_230919-a0875e5c.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101-d8_fp16_512x1024_80k_cityscapes/pspnet_r101-d8_fp16_512x1024_80k_cityscapes_20200717_230919.log.json)                             |
 | PSPNet | R-18-D8   | 769x769   |   80000 | 1.9      | 6.20           | 75.90 |         77.86 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r18-d8_769x769_80k_cityscapes.py)    | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r18-d8_769x769_80k_cityscapes/pspnet_r18-d8_769x769_80k_cityscapes_20201225_021458-3deefc62.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r18-d8_769x769_80k_cityscapes/pspnet_r18-d8_769x769_80k_cityscapes-20201225_021458.log.json)             |
@@ -55,6 +65,10 @@ Scene parsing is challenging for unrestricted open vocabulary and diverse scenes
 | PSPNet | R-18b-D8  | 769x769   |   80000 | 1.7      | 6.41           | 74.92 |         76.90 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r18b-d8_769x769_80k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r18b-d8_769x769_80k_cityscapes/pspnet_r18b-d8_769x769_80k_cityscapes_20201226_080942-bf98d186.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r18b-d8_769x769_80k_cityscapes/pspnet_r18b-d8_769x769_80k_cityscapes-20201226_080942.log.json)         |
 | PSPNet | R-50b-D8  | 769x769   |   80000 | 6.8      | 1.88           | 78.50 |         79.96 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r50b-d8_769x769_80k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50b-d8_769x769_80k_cityscapes/pspnet_r50b-d8_769x769_80k_cityscapes_20201225_094316-4c643cf6.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50b-d8_769x769_80k_cityscapes/pspnet_r50b-d8_769x769_80k_cityscapes-20201225_094316.log.json)         |
 | PSPNet | R-101b-D8 | 769x769   |   80000 | 10.8     | 1.17           | 78.87 |         80.04 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r101b-d8_769x769_80k_cityscapes.py)  | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101b-d8_769x769_80k_cityscapes/pspnet_r101b-d8_769x769_80k_cityscapes_20201226_171823-f0e7c293.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101b-d8_769x769_80k_cityscapes/pspnet_r101b-d8_769x769_80k_cityscapes-20201226_171823.log.json)     |
+| PSPNet | R-50-D32   | 512x1024  |   80000 | 3.0     | 15.21              | 73.88 |         76.85 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r50-d32_512x1024_80k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d32_512x1024_80k_cityscapes/pspnet_r50-d32_512x1024_80k_cityscapes_20220316_224840-9092b254.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d32_512x1024_80k_cityscapes/pspnet_r50-d32_512x1024_80k_cityscapes_20220316_224840.log.json)         |
+| PSPNet | R-50-D32 rsb   | 512x1024  |   80000 | 3.1 | 16.08              | 74.09 |         77.18 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes_20220316_141229-dd9c9610.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes_20220316_141229.log.json)         |
+| PSPNet | R-50b-D32   | 512x1024  |   80000 | 2.9    | 15.41              | 72.61 |         75.51 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/pspnet/pspnet_r50b-d32_512x1024_80k_cityscapes.py)   | [model](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50b-d32_512x1024_80k_cityscapes/pspnet_r50b-d32_512x1024_80k_cityscapes_20220311_152152-23bcaf8c.pth) | [log](https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50b-d32_512x1024_80k_cityscapes/pspnet_r50b-d32_512x1024_80k_cityscapes_20220311_152152.log.json)         |
+
 
 ### ADE20K
 
@@ -159,3 +173,5 @@ Note:
 
 - `FP16` means Mixed Precision (FP16) is adopted in training.
 - `896x896` is the Crop Size of iSAID dataset, which is followed by the implementation of [PointFlow: Flowing Semantics Through Points for Aerial Image Segmentation](https://arxiv.org/pdf/2103.06564.pdf)
+- `rsb` is short for 'Resnet strikes back'.
+- The `b` in `R-50b` means ResNetV1b, which is a standard ResNet backbone. In MMSegmentation, default backbone is ResNetV1c, which usually performs better in semantic segmentation task.

configs/pspnet/pspnet.yml

@@ -148,6 +148,28 @@ Models:
       mIoU(ms+flip): 79.79
   Config: configs/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes.py
   Weights: https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes/pspnet_r50-d8_512x1024_80k_cityscapes_20200606_112131-2376f12b.pth
+- Name: pspnet_r50-d8_rsb-pretrain_512x1024_adamw_80k_cityscapes
+  In Collection: PSPNet
+  Metadata:
+    backbone: R-50-D8 rsb
+    crop size: (512,1024)
+    lr schd: 80000
+    inference time (ms/im):
+    - value: 261.78
+      hardware: V100
+      backend: PyTorch
+      batch size: 1
+      mode: FP32
+      resolution: (512,1024)
+    Training Memory (GB): 6.2
+  Results:
+  - Task: Semantic Segmentation
+    Dataset: Cityscapes
+    Metrics:
+      mIoU: 78.47
+      mIoU(ms+flip): 79.45
+  Config: configs/pspnet/pspnet_r50-d8_rsb-pretrain_512x1024_adamw_80k_cityscapes.py
+  Weights: https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes/pspnet_r50-d8_rsb-pretrain_512x1024_adamw_80k_cityscapes_20220315_123238-588c30be.pth
 - Name: pspnet_r101-d8_512x1024_80k_cityscapes
   In Collection: PSPNet
   Metadata:
@@ -365,6 +387,72 @@ Models:
       mIoU(ms+flip): 80.04
   Config: configs/pspnet/pspnet_r101b-d8_769x769_80k_cityscapes.py
   Weights: https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r101b-d8_769x769_80k_cityscapes/pspnet_r101b-d8_769x769_80k_cityscapes_20201226_171823-f0e7c293.pth
+- Name: pspnet_r50-d32_512x1024_80k_cityscapes
+  In Collection: PSPNet
+  Metadata:
+    backbone: R-50-D32
+    crop size: (512,1024)
+    lr schd: 80000
+    inference time (ms/im):
+    - value: 65.75
+      hardware: V100
+      backend: PyTorch
+      batch size: 1
+      mode: FP32
+      resolution: (512,1024)
+    Training Memory (GB): 3.0
+  Results:
+  - Task: Semantic Segmentation
+    Dataset: Cityscapes
+    Metrics:
+      mIoU: 73.88
+      mIoU(ms+flip): 76.85
+  Config: configs/pspnet/pspnet_r50-d32_512x1024_80k_cityscapes.py
+  Weights: https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d32_512x1024_80k_cityscapes/pspnet_r50-d32_512x1024_80k_cityscapes_20220316_224840-9092b254.pth
+- Name: pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes
+  In Collection: PSPNet
+  Metadata:
+    backbone: R-50-D32 rsb
+    crop size: (512,1024)
+    lr schd: 80000
+    inference time (ms/im):
+    - value: 62.19
+      hardware: V100
+      backend: PyTorch
+      batch size: 1
+      mode: FP32
+      resolution: (512,1024)
+    Training Memory (GB): 3.1
+  Results:
+  - Task: Semantic Segmentation
+    Dataset: Cityscapes
+    Metrics:
+      mIoU: 74.09
+      mIoU(ms+flip): 77.18
+  Config: configs/pspnet/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes.py
+  Weights: https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes_20220316_141229-dd9c9610.pth
+- Name: pspnet_r50b-d32_512x1024_80k_cityscapes
+  In Collection: PSPNet
+  Metadata:
+    backbone: R-50b-D32
+    crop size: (512,1024)
+    lr schd: 80000
+    inference time (ms/im):
+    - value: 64.89
+      hardware: V100
+      backend: PyTorch
+      batch size: 1
+      mode: FP32
+      resolution: (512,1024)
+    Training Memory (GB): 2.9
+  Results:
+  - Task: Semantic Segmentation
+    Dataset: Cityscapes
+    Metrics:
+      mIoU: 72.61
+      mIoU(ms+flip): 75.51
+  Config: configs/pspnet/pspnet_r50b-d32_512x1024_80k_cityscapes.py
+  Weights: https://download.openmmlab.com/mmsegmentation/v0.5/pspnet/pspnet_r50b-d32_512x1024_80k_cityscapes/pspnet_r50b-d32_512x1024_80k_cityscapes_20220311_152152-23bcaf8c.pth
 - Name: pspnet_r50-d8_512x512_80k_ade20k
   In Collection: PSPNet
   Metadata:

configs/pspnet/pspnet_r50-d32_512x1024_80k_cityscapes.py

@@ -0,0 +1,5 @@
+_base_ = [
+    '../_base_/models/pspnet_r50-d8.py', '../_base_/datasets/cityscapes.py',
+    '../_base_/default_runtime.py', '../_base_/schedules/schedule_80k.py'
+]
+model = dict(backbone=dict(dilations=(1, 1, 2, 4), strides=(1, 2, 2, 2)))

configs/pspnet/pspnet_r50-d32_rsb-pretrain_512x1024_adamw_80k_cityscapes.py

@@ -0,0 +1,25 @@
+_base_ = [
+    '../_base_/models/pspnet_r50-d8.py', '../_base_/datasets/cityscapes.py',
+    '../_base_/default_runtime.py', '../_base_/schedules/schedule_80k.py'
+]
+checkpoint = 'https://download.openmmlab.com/mmclassification/v0/resnet/resnet50_8xb256-rsb-a1-600e_in1k_20211228-20e21305.pth'  # noqa
+model = dict(
+    pretrained=None,
+    backbone=dict(
+        type='ResNet',
+        init_cfg=dict(
+            type='Pretrained', prefix='backbone.', checkpoint=checkpoint),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 2, 2)))
+
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0005, weight_decay=0.05)
+optimizer_config = dict(grad_clip=dict(max_norm=1, norm_type=2))
+# learning policy
+lr_config = dict(
+    _delete_=True,
+    policy='step',
+    warmup='linear',
+    warmup_iters=1000,
+    warmup_ratio=0.001,
+    step=[60000, 72000],
+    by_epoch=False)

configs/pspnet/pspnet_r50-d8_rsb-pretrain_512x1024_adamw_80k_cityscapes.py

@@ -0,0 +1,23 @@
+_base_ = [
+    '../_base_/models/pspnet_r50-d8.py', '../_base_/datasets/cityscapes.py',
+    '../_base_/default_runtime.py', '../_base_/schedules/schedule_80k.py'
+]
+checkpoint = 'https://download.openmmlab.com/mmclassification/v0/resnet/resnet50_8xb256-rsb-a1-600e_in1k_20211228-20e21305.pth'  # noqa
+model = dict(
+    pretrained=None,
+    backbone=dict(
+        type='ResNet',
+        init_cfg=dict(
+            type='Pretrained', prefix='backbone.', checkpoint=checkpoint)))
+
+optimizer = dict(_delete_=True, type='AdamW', lr=0.0005, weight_decay=0.05)
+optimizer_config = dict(grad_clip=dict(max_norm=1, norm_type=2))
+# learning policy
+lr_config = dict(
+    _delete_=True,
+    policy='step',
+    warmup='linear',
+    warmup_iters=1000,
+    warmup_ratio=0.001,
+    step=[60000, 72000],
+    by_epoch=False)

configs/pspnet/pspnet_r50b-d32_512x1024_80k_cityscapes.py

@@ -0,0 +1,7 @@
+_base_ = [
+    '../_base_/models/pspnet_r50-d8.py', '../_base_/datasets/cityscapes.py',
+    '../_base_/default_runtime.py', '../_base_/schedules/schedule_80k.py'
+]
+model = dict(
+    pretrained='torchvision://resnet50',
+    backbone=dict(type='ResNet', dilations=(1, 1, 2, 4), strides=(1, 2, 2, 2)))

mmseg/core/evaluation/class_names.py

@@ -120,10 +120,12 @@ def isaid_classes():
         'Soccer_ball_field', 'plane', 'Harbor'
     ]
 
+
 def stare_classes():
     """stare class names for external use."""
     return ['background', 'vessel']
 
+
 def cityscapes_palette():
     """Cityscapes palette for external use."""
     return [[128, 64, 128], [244, 35, 232], [70, 70, 70], [102, 102, 156],
@@ -257,10 +259,12 @@ def isaid_palette():
             [0, 0, 191], [0, 0, 255], [0, 191, 127], [0, 127, 191],
             [0, 127, 255], [0, 100, 155]]
 
+
 def stare_palette():
     """STARE palette for external use."""
     return [[120, 120, 120], [6, 230, 230]]
 
+
 dataset_aliases = {
     'cityscapes': ['cityscapes'],
     'ade': ['ade', 'ade20k'],
@@ -274,7 +278,7 @@ def stare_palette():
         'coco_stuff164k'
     ],
     'isaid': ['isaid', 'iSAID'],
-    'stare':['stare', 'STARE']
+    'stare': ['stare', 'STARE']
 }