[Enhance] update SimCLR models and results (#295)

* [Enhance] update simclr models and results

* [Fix] revise comments to indicate settings

configs/benchmarks/classification/imagenet/resnet50_linear-8xb32-coslr-100e_in1k.py

@@ -4,10 +4,10 @@
     '../_base_/schedules/sgd_coslr-100e.py',
     '../_base_/default_runtime.py',
 ]
+# SwAV linear evaluation setting
 
 model = dict(backbone=dict(frozen_stages=4))
 
-# swav setting
 # runtime settings
 # the max_keep_ckpts controls the max number of ckpt file in your work_dirs
 # if it is 3, when CheckpointHook (in mmcv) saves the 4th ckpt

configs/benchmarks/classification/imagenet/resnet50_linear-8xb32-steplr-100e_in1k.py

@@ -4,12 +4,12 @@
     '../_base_/schedules/sgd_steplr-100e.py',
     '../_base_/default_runtime.py',
 ]
+# MoCo v1/v2 linear evaluation setting
 
 model = dict(backbone=dict(frozen_stages=4))
 
 evaluation = dict(interval=1, topk=(1, 5))
 
-# moco setting
 # optimizer
 optimizer = dict(type='SGD', lr=30., momentum=0.9, weight_decay=0.)
 

configs/benchmarks/classification/imagenet/resnet50_linear-8xb512-coslr-90e_in1k.py

@@ -4,13 +4,17 @@
     '../_base_/schedules/lars_coslr-90e.py',
     '../_base_/default_runtime.py',
 ]
+# SimSiam linear evaluation setting
+# According to SimSiam paper, this setting can also be used to evaluate
+# other methods like SimCLR, MoCo, BYOL, SwAV
 
 model = dict(backbone=dict(frozen_stages=4))
 
 # dataset summary
-data = dict(samples_per_gpu=512)  # total 512*8=4096, 8GPU linear cls
+data = dict(
+    samples_per_gpu=512,
+    workers_per_gpu=8)  # total 512*8=4096, 8GPU linear cls
 
-# simsiam setting
 # runtime settings
 # the max_keep_ckpts controls the max number of ckpt file in your work_dirs
 # if it is 3, when CheckpointHook (in mmcv) saves the 4th ckpt

configs/benchmarks/classification/imagenet/resnet50_mhead_linear-8xb32-steplr-90e_in1k.py

@@ -4,6 +4,7 @@
     '../_base_/schedules/sgd_steplr-100e.py',
     '../_base_/default_runtime.py',
 ]
+# Multi-head linear evaluation setting
 
 model = dict(backbone=dict(frozen_stages=4))
 
@@ -45,4 +46,8 @@
 
 # runtime settings
 runner = dict(type='EpochBasedRunner', max_epochs=90)
-checkpoint_config = dict(interval=10)
+
+# the max_keep_ckpts controls the max number of ckpt file in your work_dirs
+# if it is 3, when CheckpointHook (in mmcv) saves the 4th ckpt
+# it will remove the oldest one to keep the number of total ckpts as 3
+checkpoint_config = dict(interval=10, max_keep_ckpts=3)

configs/selfsup/_base_/models/simclr.py

@@ -6,7 +6,8 @@
         depth=50,
         in_channels=3,
         out_indices=[4],  # 0: conv-1, x: stage-x
-        norm_cfg=dict(type='SyncBN')),
+        norm_cfg=dict(type='SyncBN'),
+        zero_init_residual=True),
     neck=dict(
         type='NonLinearNeck',  # SimCLR non-linear neck
         in_channels=2048,

configs/selfsup/simclr/README.md

@@ -36,11 +36,12 @@ Besides, k=1 to 96 indicates the hyper-parameter of Low-shot SVM.
 
 The **Feature1 - Feature5** don't have the GlobalAveragePooling, the feature map is pooled to the specific dimensions and then follows a Linear layer to do the classification. Please refer to [resnet50_mhead_linear-8xb32-steplr-90e_in1k](https://github.com/open-mmlab/mmselfsup/blob/master/configs/benchmarks/classification/imagenet/resnet50_mhead_linear-8xb32-steplr-90e_in1k.py) for details of config.
 
-The **AvgPool** result is obtained from Linear Evaluation with GlobalAveragePooling. Please refer to [resnet50_linear-8xb32-steplr-100e_in1k](https://github.com/open-mmlab/mmselfsup/blob/master/configs/benchmarks/classification/imagenet/resnet50_linear-8xb32-steplr-100e_in1k.py) for details of config.
+The **AvgPool** result is obtained from Linear Evaluation with GlobalAveragePooling. Please refer to [resnet50_linear-8xb512-coslr-90e_in1k](https://github.com/open-mmlab/mmselfsup/blob/master/configs/benchmarks/classification/imagenet/resnet50_linear-8xb512-coslr-90e_in1k.py) for details of config.
 
-| Self-Supervised Config                                                                                                                           | Feature1 | Feature2 | Feature3 | Feature4 | Feature5 | AvgPool |
-| ------------------------------------------------------------------------------------------------------------------------------------------------ | -------- | -------- | -------- | -------- | -------- | ------- |
-| [resnet50_8xb32-coslr-200e](https://github.com/open-mmlab/mmselfsup/blob/master/configs/selfsup/simclr/simclr_resnet50_8xb32-coslr-200e_in1k.py) | 14.43    | 30.97    | 41.02    | 53.92    | 61.24    | 57.28   |
+| Self-Supervised Config                                                                                                                               | Feature1 | Feature2 | Feature3 | Feature4 | Feature5 | AvgPool |
+| ---------------------------------------------------------------------------------------------------------------------------------------------------- | -------- | -------- | -------- | -------- | -------- | ------- |
+| [resnet50_8xb32-coslr-200e](https://github.com/open-mmlab/mmselfsup/blob/master/configs/selfsup/simclr/simclr_resnet50_8xb32-coslr-200e_in1k.py)     | 16.29    | 31.11    | 39.99    | 55.06    | 62.91    | 62.56   |
+| [resnet50_16xb256-coslr-200e](https://github.com/open-mmlab/mmselfsup/blob/master/configs/selfsup/simclr/simclr_resnet50_16xb256-coslr-200e_in1k.py) | 15.44    | 31.47    | 41.83    | 59.44    | 66.41    | 66.66   |
 
 #### Places205 Linear Evaluation
 

configs/selfsup/simclr/metafile.yml

@@ -4,7 +4,7 @@ Collections:
       Training Data: ImageNet-1k
       Training Techniques:
         - LARS
-      Training Resources: 8x V100 GPUs
+      Training Resources: 8x V100 GPUs (b256), 16x A100-80G GPUs (b4096)
       Architecture:
         - ResNet
         - SimCLR
@@ -23,6 +23,18 @@ Models:
       - Task: Self-Supervised Image Classification
         Dataset: ImageNet-1k
         Metrics:
-          Top 1 Accuracy: 57.28
+          Top 1 Accuracy: 62.56
     Config: configs/selfsup/simclr/simclr_resnet50_8xb32-coslr-200e_in1k.py
-    Weights: https://download.openmmlab.com/mmselfsup/simclr/simclr_resnet50_8xb32-coslr-200e_in1k_20220225-97d2abef.pth
+    Weights: https://download.openmmlab.com/mmselfsup/simclr/simclr_resnet50_8xb32-coslr-200e_in1k_20220428-46ef6bb9.pth
+  - Name: simclr_resnet50_16xb256-coslr-200e_in1k
+    In Collection: SimCLR
+    Metadata:
+      Epochs: 200
+      Batch Size: 4096
+    Results:
+      - Task: Self-Supervised Image Classification
+        Dataset: ImageNet-1k
+        Metrics:
+          Top 1 Accuracy: 66.66
+    Config: configs/selfsup/simclr/simclr_resnet50_16xb256-coslr-200e_in1k.py
+    Weights: https://download.openmmlab.com/mmselfsup/simclr/simclr_resnet50_16xb256-coslr-200e_in1k_20220428-8c24b063.pth

configs/selfsup/simclr/simclr_resnet50_16xb256-coslr-200e_in1k.py

@@ -0,0 +1,7 @@
+_base_ = 'simclr_resnet50_8xb32-coslr-200e_in1k.py'
+
+# optimizer
+optimizer = dict(lr=4.8)
+
+# dataset summary
+data = dict(samples_per_gpu=256, workers_per_gpu=8)  # total 256*16

configs/selfsup/simclr/simclr_resnet50_8xb64-coslr-200e_in1k.py

@@ -1,4 +1,7 @@
 _base_ = 'simclr_resnet50_8xb32-coslr-200e_in1k.py'
 
+# optimizer
+optimizer = dict(lr=0.6)
+
 # dataset summary
 data = dict(samples_per_gpu=64)  # total 64*8