Merge branch 'open-mmlab:master' into fix_segmentor_inference

configs/bisenetv1/bisenetv1_r101-d32_lr5e-3_4x4_512x512_160k_coco-stuff164k.py

@@ -3,6 +3,7 @@
  '../_base_/datasets/coco-stuff164k.py', '../_base_/default_runtime.py',
  '../_base_/schedules/schedule_160k.py'
 ]
+norm_cfg = dict(type='SyncBN', requires_grad=True)
 model = dict(
  backbone=dict(
  context_channels=(512, 1024, 2048),
@@ -11,8 +12,30 @@
  backbone_cfg=dict(type='ResNet', depth=101)),
  decode_head=dict(in_channels=1024, channels=1024, num_classes=171),
  auxiliary_head=[
- dict(in_channels=512, channels=256, num_classes=171),
- dict(in_channels=512, channels=256, num_classes=171),
+ dict(
+ type='FCNHead',
+ in_channels=512,
+ channels=256,
+ num_convs=1,
+ num_classes=171,
+ in_index=1,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+ dict(
+ type='FCNHead',
+ in_channels=512,
+ channels=256,
+ num_convs=1,
+ num_classes=171,
+ in_index=2,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
  ])
 lr_config = dict(warmup='linear', warmup_iters=1000)
 optimizer = dict(lr=0.005)

configs/bisenetv1/bisenetv1_r18-d32_lr5e-3_4x4_512x512_160k_coco-stuff164k.py

@@ -3,11 +3,34 @@
  '../_base_/datasets/coco-stuff164k.py', '../_base_/default_runtime.py',
  '../_base_/schedules/schedule_160k.py'
 ]
+norm_cfg = dict(type='SyncBN', requires_grad=True)
 model = dict(
  decode_head=dict(num_classes=171),
  auxiliary_head=[
- dict(num_classes=171),
- dict(num_classes=171),
+ dict(
+ type='FCNHead',
+ in_channels=128,
+ channels=64,
+ num_convs=1,
+ num_classes=171,
+ in_index=1,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+ dict(
+ type='FCNHead',
+ in_channels=128,
+ channels=64,
+ num_convs=1,
+ num_classes=171,
+ in_index=2,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
  ])
 lr_config = dict(warmup='linear', warmup_iters=1000)
 optimizer = dict(lr=0.005)

configs/bisenetv1/bisenetv1_r50-d32_lr5e-3_4x4_512x512_160k_coco-stuff164k.py

@@ -3,6 +3,7 @@
  '../_base_/datasets/coco-stuff164k.py', '../_base_/default_runtime.py',
  '../_base_/schedules/schedule_160k.py'
 ]
+norm_cfg = dict(type='SyncBN', requires_grad=True)
 model = dict(
  backbone=dict(
  context_channels=(512, 1024, 2048),
@@ -11,8 +12,30 @@
  backbone_cfg=dict(type='ResNet', depth=50)),
  decode_head=dict(in_channels=1024, channels=1024, num_classes=171),
  auxiliary_head=[
- dict(in_channels=512, channels=256, num_classes=171),
- dict(in_channels=512, channels=256, num_classes=171),
+ dict(
+ type='FCNHead',
+ in_channels=512,
+ channels=256,
+ num_convs=1,
+ num_classes=171,
+ in_index=1,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+ dict(
+ type='FCNHead',
+ in_channels=512,
+ channels=256,
+ num_convs=1,
+ num_classes=171,
+ in_index=2,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
  ])
 lr_config = dict(warmup='linear', warmup_iters=1000)
 optimizer = dict(lr=0.005)

configs/bisenetv2/README.md

@@ -42,7 +42,7 @@ The low-level details and high-level semantics are both essential to the semanti
 | Method | Backbone | Crop Size | Lr schd | Mem (GB) | Inf time (fps) | mIoU | mIoU(ms+flip) | config | download |
 | ---------------- | --------- | --------- | ------: | -------- | -------------- | ----: | ------------: | ---------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
 | BiSeNetV2 | BiSeNetV2 | 1024x1024 | 160000 | 7.64 | 31.77 | 73.21 | 75.74 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes_20210902_015551-bcf10f09.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_4x4_1024x1024_160k_cityscapes_20210902_015551.log.json) |
-| BiSeNetV2 (OHEM) | BiSeNetV2 | 1024x1024 | 160000 | 7.64 | - | 73.57 | 75.80 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20210902_112947-5f8103b4.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20210902_112947.log.json) |
+| BiSeNetV2 (OHEM) | BiSeNetV2 | 1024x1024 | 160000 | 7.64 | - | 75.30 | 77.06 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20220808_172324-8bf0aaba.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20220808_172324.log.json) |
 | BiSeNetV2 (4x8) | BiSeNetV2 | 1024x1024 | 160000 | 15.05 | - | 75.76 | 77.79 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes_20210903_000032-e1a2eed6.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes/bisenetv2_fcn_4x8_1024x1024_160k_cityscapes_20210903_000032.log.json) |
 | BiSeNetV2 (FP16) | BiSeNetV2 | 1024x1024 | 160000 | 5.77 | 36.65 | 73.07 | 75.13 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes_20210902_045942-b979777b.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes_20210902_045942.log.json) |
 

configs/bisenetv2/bisenetv2.yml

@@ -45,10 +45,10 @@ Models:
  - Task: Semantic Segmentation
  Dataset: Cityscapes
  Metrics:
- mIoU: 73.57
- mIoU(ms+flip): 75.8
+ mIoU: 75.3
+ mIoU(ms+flip): 77.06
  Config: configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py
- Weights: https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20210902_112947-5f8103b4.pth
+ Weights: https://download.openmmlab.com/mmsegmentation/v0.5/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes_20220808_172324-8bf0aaba.pth
 - Name: bisenetv2_fcn_4x8_1024x1024_160k_cityscapes
  In Collection: BiSeNetV2
  Metadata:

configs/bisenetv2/bisenetv2_fcn_ohem_4x4_1024x1024_160k_cityscapes.py

@@ -3,7 +3,67 @@
  '../_base_/datasets/cityscapes_1024x1024.py',
  '../_base_/default_runtime.py', '../_base_/schedules/schedule_160k.py'
 ]
-sampler = dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000)
+# sampler = dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000)
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+ decode_head=dict(
+ sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000)),
+ auxiliary_head=[
+ dict(
+ type='FCNHead',
+ in_channels=16,
+ channels=16,
+ num_convs=2,
+ num_classes=19,
+ in_index=1,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000),
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+ dict(
+ type='FCNHead',
+ in_channels=32,
+ channels=64,
+ num_convs=2,
+ num_classes=19,
+ in_index=2,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000),
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+ dict(
+ type='FCNHead',
+ in_channels=64,
+ channels=256,
+ num_convs=2,
+ num_classes=19,
+ in_index=3,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000),
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+ dict(
+ type='FCNHead',
+ in_channels=128,
+ channels=1024,
+ num_convs=2,
+ num_classes=19,
+ in_index=4,
+ norm_cfg=norm_cfg,
+ concat_input=False,
+ align_corners=False,
+ sampler=dict(type='OHEMPixelSampler', thresh=0.7, min_kept=10000),
+ loss_decode=dict(
+ type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+ ],
+)
+
 lr_config = dict(warmup='linear', warmup_iters=1000)
 optimizer = dict(lr=0.05)
 data = dict(

configs/knet/README.md

@@ -43,7 +43,7 @@ Semantic, instance, and panoptic segmentations have been addressed using differe
 | KNet + UPerNet | R-50-D8 | 512x512 | 80000 | 7.34 | 17.11 | 43.45 | 44.07 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/knet/knet_s3_upernet_r50-d8_8x2_512x512_adamw_80k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_r50-d8_8x2_512x512_adamw_80k_ade20k/knet_s3_upernet_r50-d8_8x2_512x512_adamw_80k_ade20k_20220304_125657-215753b0.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_r50-d8_8x2_512x512_adamw_80k_ade20k/knet_s3_upernet_r50-d8_8x2_512x512_adamw_80k_ade20k_20220304_125657.log.json) |
 | KNet + UPerNet | Swin-T | 512x512 | 80000 | 7.57 | 15.56 | 45.84 | 46.27 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/knet/knet_s3_upernet_swin-t_8x2_512x512_adamw_80k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-t_8x2_512x512_adamw_80k_ade20k/knet_s3_upernet_swin-t_8x2_512x512_adamw_80k_ade20k_20220303_133059-7545e1dc.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-t_8x2_512x512_adamw_80k_ade20k/knet_s3_upernet_swin-t_8x2_512x512_adamw_80k_ade20k_20220303_133059.log.json) |
 | KNet + UPerNet | Swin-L | 512x512 | 80000 | 13.5 | 8.29 | 52.05 | 53.24 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/knet/knet_s3_upernet_swin-l_8x2_512x512_adamw_80k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-l_8x2_512x512_adamw_80k_ade20k/knet_s3_upernet_swin-l_8x2_512x512_adamw_80k_ade20k_20220303_154559-d8da9a90.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-l_8x2_512x512_adamw_80k_ade20k/knet_s3_upernet_swin-l_8x2_512x512_adamw_80k_ade20k_20220303_154559.log.json) |
-| KNet + UPerNet | Swin-L | 640x640 | 80000 | 13.54 | 8.29 | 52.21 | 53.34 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k_20220301_220747-8787fc71.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k_20220301_220747.log.json) |
+| KNet + UPerNet | Swin-L | 640x640 | 80000 | 18.31 | 5.55 | 52.46 | 53.78 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k_20220720_165636-cbcaed32.pth) \| [log](https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k_20220720_165636.log.json) |
 
 Note:
 

configs/knet/knet.yml

@@ -152,18 +152,18 @@ Models:
  crop size: (640,640)
  lr schd: 80000
  inference time (ms/im):
- - value: 120.63
+ - value: 180.18
  hardware: V100
  backend: PyTorch
  batch size: 1
  mode: FP32
  resolution: (640,640)
- Training Memory (GB): 13.54
+ Training Memory (GB): 18.31
  Results:
  - Task: Semantic Segmentation
  Dataset: ADE20K
  Metrics:
- mIoU: 52.21
- mIoU(ms+flip): 53.34
+ mIoU: 52.46
+ mIoU(ms+flip): 53.78
  Config: configs/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k.py
- Weights: https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k_20220301_220747-8787fc71.pth
+ Weights: https://download.openmmlab.com/mmsegmentation/v0.5/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k_20220720_165636-cbcaed32.pth

configs/knet/knet_s3_upernet_swin-l_8x2_640x640_adamw_80k_ade20k.py

@@ -46,9 +46,10 @@
  dict(type='Collect', keys=['img']),
  ])
 ]
+# In K-Net implementation we use batch size 2 per GPU as default
 data = dict(
+ samples_per_gpu=2,
+ workers_per_gpu=2,
  train=dict(pipeline=train_pipeline),
  val=dict(pipeline=test_pipeline),
  test=dict(pipeline=test_pipeline))
-# In K-Net implementation we use batch size 2 per GPU as default
-data = dict(samples_per_gpu=2, workers_per_gpu=2)