Update for mmdet v2.25.0

WonderNoot · Jun 18, 2022 · a8661f8 · a8661f8
1 parent 13d2903
commit a8661f8
Show file tree

Hide file tree

Showing 25 changed files with 150 additions and 93 deletions.
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
@@ -28,16 +28,15 @@ repos:
     rev: v2.1.0
     hooks:
       - id: codespell
-  # TODO
-  # - repo: https://github.com/executablebooks/mdformat
-  #   rev: 0.7.14
-  #   hooks:
-  #     - id: mdformat
-  #       args: ["--number"]
-  #       additional_dependencies:
-  #         - mdformat-gfm
-  #         - mdformat_frontmatter
-  #         - linkify-it-py
+  - repo: https://github.com/executablebooks/mdformat
+    rev: 0.7.14
+    hooks:
+      - id: mdformat
+        args: ["--number"]
+        additional_dependencies:
+          - mdformat-gfm
+          - mdformat_frontmatter
+          - linkify-it-py
   - repo: https://github.com/myint/docformatter
     rev: v1.3.1
     hooks:

diff --git a/README.md b/README.md
@@ -11,7 +11,7 @@ https://arxiv.org/abs/2103.14027
 
 - recent:
   - Add FocalNet, GBR COTS dataset
-  - Update codes for mmdet 2.22.0, mmcv-full 1.4.4
+  - Update codes for mmdet 2.25.0, mmcv-full 1.4.4
 - 21.12 (Dec. 2021):
   - Support [finer scale-wise AP metrics](docs/en/tutorials/finer_scale_ap.md)
   - Add codes for TOOD, ConvMLP, PoolFormer
@@ -136,7 +136,6 @@ When merging, please note that there are some minor differences from the above r
 
 <br><br>
 
-
 <div align="center">
   <img src="resources/mmdet-logo.png" width="600"/>
   <div>&nbsp;</div>

diff --git a/configs/cbnet/README.md b/configs/cbnet/README.md
diff --git a/...s/cbnet/cascade_mask_rcnn_cbv2_swin_base_patch4_window7_mstrain_400-1400_adamw_3x_coco.py b/...s/cbnet/cascade_mask_rcnn_cbv2_swin_base_patch4_window7_mstrain_400-1400_adamw_3x_coco.py
@@ -44,4 +44,9 @@
     train=dict(pipeline=train_pipeline),
     val=dict(pipeline=test_pipeline),
     test=dict(pipeline=test_pipeline))
-optimizer = dict(lr=0.0001 * (samples_per_gpu / 2))
+optimizer = dict(lr=0.0001 / 2)
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR,
+# USER SHOULD NOT CHANGE ITS VALUES.
+# base_batch_size = (8 GPUs) x (1 samples per GPU)
+auto_scale_lr = dict(base_batch_size=8)
diff --git a/.../cbnet/cascade_mask_rcnn_cbv2_swin_small_patch4_window7_mstrain_400-1400_adamw_3x_coco.py b/.../cbnet/cascade_mask_rcnn_cbv2_swin_small_patch4_window7_mstrain_400-1400_adamw_3x_coco.py
@@ -44,4 +44,9 @@
     train=dict(pipeline=train_pipeline),
     val=dict(pipeline=test_pipeline),
     test=dict(pipeline=test_pipeline))
-optimizer = dict(lr=0.0001 * (samples_per_gpu / 2))
+optimizer = dict(lr=0.0001 / 2)
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR,
+# USER SHOULD NOT CHANGE ITS VALUES.
+# base_batch_size = (8 GPUs) x (1 samples per GPU)
+auto_scale_lr = dict(base_batch_size=8)
diff --git a/configs/cbnet/htc_cbv2_swin_base_patch4_window7_mstrain_400-1400_adamw_20e_coco.py b/configs/cbnet/htc_cbv2_swin_base_patch4_window7_mstrain_400-1400_adamw_20e_coco.py
@@ -85,7 +85,7 @@
 optimizer = dict(
     _delete_=True,
     type='AdamW',
-    lr=0.0001 * (samples_per_gpu / 2),
+    lr=0.0001 / 2,
     betas=(0.9, 0.999),
     weight_decay=0.05,
     paramwise_cfg=dict(
@@ -96,3 +96,8 @@
         }))
 
 fp16 = dict(loss_scale=dict(init_scale=512))
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR,
+# USER SHOULD NOT CHANGE ITS VALUES.
+# base_batch_size = (8 GPUs) x (1 samples per GPU)
+auto_scale_lr = dict(base_batch_size=8)
diff --git a/configs/cbnet/model_zoo.md b/configs/cbnet/model_zoo.md
@@ -9,18 +9,21 @@
 | DB-ResNet50 |   1x    |       40.8        |   69M   | 284G  | [config](faster_rcnn_cbv2d1_r50_fpn_1x_coco.py) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/faster_rcnn_cbv2d1_r50_fpn_1x_coco.log.json) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/faster_rcnn_cbv2d1_r50_fpn_1x_coco.pth.zip) |
 
 ### Cascade R-CNN (1600x1400)
+
 |     Backbone      |    Lr Schd     | box mAP (minival/test-dev) | #params | FLOPs |                                         config                                          |                                                                       model                                                                        |
 | :---------------: | :------------: | :------------------------: | :-----: | :---: | :-------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------------------------: |
 | DB-Res2Net101-DCN |      20e       |           53.7/-           |  141M   | 429G  |       [config](cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_coco.py)       |   [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_coco.pth.zip)   |
 | DB-Res2Net101-DCN | 20e + 1x (swa) |         54.8/55.3          |  141M   | 429G  | [config (test only)](cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_coco.py) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_coco_swa.pth.zip) |
 
 ### Cascade R-CNN w/ 4conv1fc (1600x1400)
+
 |     Backbone      |    Lr Schd     | box mAP (minival/test-dev) | #params | FLOPs |                                                config                                                |                                                                              model                                                                              |
 | :---------------: | :------------: | :------------------------: | :-----: | :---: | :--------------------------------------------------------------------------------------------------: | :-------------------------------------------------------------------------------------------------------------------------------------------------------------: |
 | DB-Res2Net101-DCN |      20e       |           54.1/-           |  146M   | 774G  |       [config](cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_giou_4conv1f_coco.py)       |   [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_giou_4conv1f_coco.pth.zip)   |
 | DB-Res2Net101-DCN | 20e + 1x (swa) |         55.3/55.6          |  146M   | 774G  | [config (test only)](cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_giou_4conv1f_coco.py) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_giou_4conv1f_coco_swa.pth.zip) |
 
 **Notes**:
+
 - For SWA training, please refer to [SWA Object Detection](https://github.com/hyz-xmaster/swa_object_detection)
 
 ## Transformer-based (w/ ImageNet-1k pretrained)
@@ -32,18 +35,21 @@
 | DB-Swin-T |   3x    |       50.2        |        44.5        |   76M   | 357G  | [config](mask_rcnn_cbv2_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/mask_rcnn_cbv2_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.log.json) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/mask_rcnn_cbv2_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.pth.zip) |
 
 ### Cascade Mask R-CNN w/ 4conv1fc
+
 | Backbone  | Lr Schd | box mAP (minival) | mask mAP (minival) | #params | FLOPs |                                           config                                           |                                                                              log                                                                               |                                                                             model                                                                             |
 | :-------: | :-----: | :---------------: | :----------------: | :-----: | :---: | :----------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------------------------------: |
 | DB-Swin-T |   3x    |       53.6        |        46.2        |  114M   | 836G  | [config](cascade_mask_rcnn_cbv2_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/cascade_mask_rcnn_cbv2_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.log.json) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/cascade_mask_rcnn_cbv2_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.pth.zip) |
 
 ### Cascade Mask R-CNN w/ 4conv1fc (1600x1400)
+
 | Backbone  | Lr Schd | box mAP (minival/test-dev) | mask mAP (minival/test-dev) | #params | FLOPs |                                            config                                            |                                                                              model                                                                              |
 | :-------: | :-----: | :------------------------: | :-------------------------: | :-----: | :---: | :------------------------------------------------------------------------------------------: | :-------------------------------------------------------------------------------------------------------------------------------------------------------------: |
 | DB-Swin-S |   3x    |         56.3/56.9          |          48.6/49.1          |  156M   | 1016G | [config](cascade_mask_rcnn_cbv2_swin_small_patch4_window7_mstrain_400-1400_adamw_3x_coco.py) | [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/cascade_mask_rcnn_cbv2_swin_small_patch4_window7_mstrain_400-1400_adamw_3x_coco.pth.zip) |
 
 ## Transformer-based (w/ ImageNet-22k pretrained)
 
 ### HTC (1600x1400)
+
 | Backbone  |    Lr Schd     | box mAP (minival/test-dev) | mask mAP (minival/test-dev) | #params | FLOPs |                                           config                                           |                                                                          model                                                                           |
 | :-------: | :------------: | :------------------------: | :-------------------------: | :-----: | :---: | :----------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------------------------------: |
 | DB-Swin-B |      20e       |           57.9/-           |           50.2/-            |  231M   | 1004G |       [config](htc_cbv2_swin_base_patch4_window7_mstrain_400-1400_adamw_20e_coco.py)       |   [github](https://github.com/CBNetwork/storage/releases/download/v1.0.0/htc_cbv2_swin_base22k_patch4_window7_mstrain_400-1400_adamw_20e_coco.pth.zip)   |

diff --git a/...gbr_cots/cascade_rcnn_r2_101_fpn_fp16_4x2_mixup_affine_hsv_1440_lr002_7e_cots_alltrain.py b/...gbr_cots/cascade_rcnn_r2_101_fpn_fp16_4x2_mixup_affine_hsv_1440_lr002_7e_cots_alltrain.py
@@ -27,3 +27,8 @@
 custom_hooks = [
     dict(type='YOLOXModeSwitchHook', num_last_epochs=0, priority=48)
 ]
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR,
+# USER SHOULD NOT CHANGE ITS VALUES.
+# base_batch_size = (4 GPUs) x (2 samples per GPU)
+auto_scale_lr = dict(base_batch_size=8)
diff --git a/...gle_gbr_cots/cascade_rcnn_r2_101_fpn_fp16_4x2_mixup_affine_hsv_1440_lr002_7e_cots_val2.py b/...gle_gbr_cots/cascade_rcnn_r2_101_fpn_fp16_4x2_mixup_affine_hsv_1440_lr002_7e_cots_val2.py
@@ -27,3 +27,8 @@
 custom_hooks = [
     dict(type='YOLOXModeSwitchHook', num_last_epochs=0, priority=48)
 ]
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR,
+# USER SHOULD NOT CHANGE ITS VALUES.
+# base_batch_size = (4 GPUs) x (2 samples per GPU)
+auto_scale_lr = dict(base_batch_size=8)
diff --git a/configs/focalnet/README.md b/configs/focalnet/README.md
@@ -21,6 +21,9 @@ See the [official repository](https://github.com/microsoft/FocalNet/tree/ecfa580
 ### Notes
 
 - We may need key conversion for the official Sparse R-CNN weights.
+- Although most configs set `samples_per_gpu=1`, the authors overwrite it with `samples_per_gpu=2`.
+  See [detection/README.md](https://github.com/microsoft/FocalNet/blob/main/detection/README.md) and `"iter": 7330` in [logs](https://github.com/microsoft/FocalNet/tree/ecfa580e252a106899e03de13af543f580f43da2#object-detection-on-coco).
+  Thus we keep `base_batch_size=16` in `auto_scale_lr`.
 - If you find performance difference from apex (used by the original authors), please raise an issue.
 
 ## Citation

diff --git a/configs/gflv2/README.md b/configs/gflv2/README.md
@@ -18,11 +18,11 @@ For details see [GFocalV2](https://arxiv.org/abs/2011.12885).
 | X-101-dcn  |   2x    |         Yes          |  48.8  |      10.7      | [model](https://drive.google.com/file/d/1AHDVQoclYPSP0Ync2a5FCsr_rhq2QdMH/view?usp=sharing) |
 | R2-101-dcn |   2x    |         Yes          |  49.9  |      10.9      | [model](https://drive.google.com/file/d/1sAXfYLXIxZgMrC44LBqDgfYImThZ_kud/view?usp=sharing) |
 
-[1] *The reported numbers here are from new experimental trials (in the cleaned repo), which may be slightly different from the original paper.* \
-[2] *Note that the 1x performance may be slightly unstable due to insufficient training. In practice, the 2x results are considerably stable between multiple runs.* \
-[3] *All results are obtained with a single model and without any test time data augmentation such as multi-scale, flipping and etc..* \
-[4] *`dcn` denotes deformable convolutional networks.* \
-[5] *FPS is tested with a single GeForce RTX 2080Ti GPU, using a batch size of 1.*
+\[1\] *The reported numbers here are from new experimental trials (in the cleaned repo), which may be slightly different from the original paper.* \
+\[2\] *Note that the 1x performance may be slightly unstable due to insufficient training. In practice, the 2x results are considerably stable between multiple runs.* \
+\[3\] *All results are obtained with a single model and without any test time data augmentation such as multi-scale, flipping and etc..* \
+\[4\] *`dcn` denotes deformable convolutional networks.* \
+\[5\] *FPS is tested with a single GeForce RTX 2080Ti GPU, using a batch size of 1.*
 
 ## Citation
 

diff --git a/configs/manga109/README.md b/configs/manga109/README.md
@@ -16,43 +16,39 @@ Please see [our manga109api fork](https://github.com/shinya7y/manga109api) to co
 We use [68train, 4val, and 15test splits](#dataset-splits).
 The 15test set was selected to be well-balanced for reliable evaluation.
 
-
 ## Results
 
 ### 68train, 15test (Manga109-s v2020.12.18)
 
-|      Method       | Backbone | Lr schd |  AP   |                                                                             Download                                                                             |
-| :---------------: | :------: | :-----: | :---: | :--------------------------------------------------------------------------------------------------------------------------------------------------------------: |
-|   Faster R-CNN    |   R-50   |   1x    | 65.8  |        [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/faster_rcnn_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-264d9f31.pth)         |
-|   Cascade R-CNN   |   R-50   |   1x    | 67.6  |        [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/cascade_rcnn_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-aece91e1.pth)        |
-|     RetinaNet     |   R-50   |   1x    | 65.3  |         [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/retinanet_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-9fa45ba4.pth)          |
-|       ATSS        |   R-50   |   1x    | 66.5  |            [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/atss_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-c3e34e96.pth)            |
-|       ATSS        |  Swin-T  |   1x    | 66.2  |         [model](https://github.com/shinya7y/weights/releases/download/v1.0.1/atss_swint_fpn_fp16_4x4_adamw_1x_manga109s_20210504_epoch_12-c96ddec3.pth)          |
-|     ATSS+SEPC     |   R-50   |   1x    | 67.1  |      [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/atss_r50_fpn_sepc_noibn_fp16_4x4_1x_manga109s_20201219_epoch_12-b9eef036.pth)       |
-|        GFL        |   R-50   |   1x    | 67.3  |            [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/gfl_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-49659797.pth)             |
-|    UniverseNet    |  R2-50   |   1x    | 68.9  |   [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/universenet50_fp16_4x4_mstrain_480_960_1x_manga109s_20201220_epoch_12-ae4e7451.pth)    |
-| UniverseNet 20.08 |  R2-50   |   1x    | 69.9  | [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/universenet50_2008_fp16_4x4_mstrain_480_960_1x_manga109s_20201220_epoch_12-6af914a4.pth) |
+|      Method       | Backbone | Lr schd |  AP  |                                                                             Download                                                                             |
+| :---------------: | :------: | :-----: | :--: | :--------------------------------------------------------------------------------------------------------------------------------------------------------------: |
+|   Faster R-CNN    |   R-50   |   1x    | 65.8 |        [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/faster_rcnn_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-264d9f31.pth)         |
+|   Cascade R-CNN   |   R-50   |   1x    | 67.6 |        [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/cascade_rcnn_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-aece91e1.pth)        |
+|     RetinaNet     |   R-50   |   1x    | 65.3 |         [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/retinanet_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-9fa45ba4.pth)          |
+|       ATSS        |   R-50   |   1x    | 66.5 |            [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/atss_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-c3e34e96.pth)            |
+|       ATSS        |  Swin-T  |   1x    | 66.2 |         [model](https://github.com/shinya7y/weights/releases/download/v1.0.1/atss_swint_fpn_fp16_4x4_adamw_1x_manga109s_20210504_epoch_12-c96ddec3.pth)          |
+|     ATSS+SEPC     |   R-50   |   1x    | 67.1 |      [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/atss_r50_fpn_sepc_noibn_fp16_4x4_1x_manga109s_20201219_epoch_12-b9eef036.pth)       |
+|        GFL        |   R-50   |   1x    | 67.3 |            [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/gfl_r50_fpn_fp16_4x4_1x_manga109s_20201219_epoch_12-49659797.pth)             |
+|    UniverseNet    |  R2-50   |   1x    | 68.9 |   [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/universenet50_fp16_4x4_mstrain_480_960_1x_manga109s_20201220_epoch_12-ae4e7451.pth)    |
+| UniverseNet 20.08 |  R2-50   |   1x    | 69.9 | [model](https://github.com/shinya7y/UniverseNet/releases/download/20.12/universenet50_2008_fp16_4x4_mstrain_480_960_1x_manga109s_20201220_epoch_12-6af914a4.pth) |
 
 - In addition to ATSS+SEPC, UniverseNet uses Res2Net-v1b-50, DCN, and multi-scale training (480-960).
 - The settings for normalization layers (including whether to use iBN of SEPC) depend on the config files.
 - All models were trained and evaluated using fp16 (mixed precision).
 - Each model was fine-tuned from a corresponding COCO pre-trained model.
 
-
 ## Dataset splits
 
-- 15test: ["Akuhamu", "BakuretsuKungFuGirl", "DollGun", "EvaLady", "HinagikuKenzan", "KyokugenCyclone", "LoveHina_vol01", "MomoyamaHaikagura", "TennenSenshiG", "UchiNoNyan'sDiary", "UnbalanceTokyo", "YamatoNoHane", "YoumaKourin", "YumeNoKayoiji", "YumeiroCooking"]
-- 4val: ["HealingPlanet", "LoveHina_vol14", "SeisinkiVulnus", "That'sIzumiko"]
+- 15test: `["Akuhamu", "BakuretsuKungFuGirl", "DollGun", "EvaLady", "HinagikuKenzan", "KyokugenCyclone", "LoveHina_vol01", "MomoyamaHaikagura", "TennenSenshiG", "UchiNoNyan'sDiary", "UnbalanceTokyo", "YamatoNoHane", "YoumaKourin", "YumeNoKayoiji", "YumeiroCooking"]`
+- 4val: `["HealingPlanet", "LoveHina_vol14", "SeisinkiVulnus", "That'sIzumiko"]`
 - 68train: All the other volumes
 
-
 ## Notes
 
 - Please check the dataset licenses ([Manga109](http://www.manga109.org/en/download.html), [Manga109-s](http://www.manga109.org/en/download_s.html)).
 - The typical scale of the original images is (1654, 1170).
   The number of maximum total pixels of (1216, 864) for Manga109 is almost the same as that of (1333, 800) for COCO.
 
-
 ## Citations
 
 Users must cite the two papers below for use in academic papers.