mirrors
/
mmselfsup
mirror of https://github.com/open-mmlab/mmselfsup.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #24 from xieenze/master 

update coco & byol results
pull/25/head
Xiaohang Zhan 2020-08-13 12:01:10 +08:00 committed by GitHub
parent c873fee832 2e921960ad
commit 9160c35c7c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 112 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

98
configs/selfsup/byol/r50_bs4096_ep200.py 100644
View File

@ -0,0 +1,98 @@
import copy
_base_ = '../../base.py'
# model settings
model = dict(
type='BYOL',
pretrained=None,
base_momentum=0.996,
backbone=dict(
type='ResNet',
depth=50,
in_channels=3,
out_indices=[4], # 0: conv-1, x: stage-x
norm_cfg=dict(type='BN')),
neck=dict(
type='NonLinearNeckV2',
in_channels=2048,
hid_channels=4096,
out_channels=256,
with_avg_pool=True),
head=dict(type='LatentPredictHead',
size_average=True,
predictor=dict(type='NonLinearNeckV2',
in_channels=256, hid_channels=4096,
out_channels=256, with_avg_pool=False)))
# dataset settings
data_source_cfg = dict(
type='ImageNet',
memcached=True,
mclient_path='/mnt/lustre/share/memcached_client')
data_train_list = 'data/imagenet/meta/train.txt'
data_train_root = 'data/imagenet/train'
dataset_type = 'BYOLDataset'
img_norm_cfg = dict(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
train_pipeline = [
dict(type='RandomResizedCrop', size=224, interpolation=3), # bicubic
dict(type='RandomHorizontalFlip'),
dict(
type='RandomAppliedTrans',
transforms=[
dict(
type='ColorJitter',
brightness=0.4,
contrast=0.4,
saturation=0.2,
hue=0.1)
],
p=0.8),
dict(type='RandomGrayscale', p=0.2),
dict(
type='RandomAppliedTrans',
transforms=[
dict(
type='GaussianBlur',
sigma_min=0.1,
sigma_max=2.0,
kernel_size=23)
],
p=1.),
dict(type='RandomAppliedTrans',
transforms=[dict(type='Solarization')], p=0.),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg),
]
train_pipeline1 = copy.deepcopy(train_pipeline)
train_pipeline2 = copy.deepcopy(train_pipeline)
train_pipeline2[4]['p'] = 0.1 # gaussian blur
train_pipeline2[5]['p'] = 0.2 # solarization
data = dict(
imgs_per_gpu=128, # total 128*32=4096(32gpus)
workers_per_gpu=4,
train=dict(
type=dataset_type,
data_source=dict(
list_file=data_train_list, root=data_train_root,
**data_source_cfg),
pipeline1=train_pipeline1,
pipeline2=train_pipeline2))
# additional hooks
custom_hooks = [
dict(type='BYOLHook', end_momentum=1.)
]
# optimizer
optimizer = dict(type='LARS', lr=3.2, weight_decay=0.0000015, momentum=0.9,
paramwise_options={
'(bn|gn)(\d+)?.(weight|bias)': dict(weight_decay=0., lars_exclude=True),
'bias': dict(weight_decay=0., lars_exclude=True)})
# learning policy
lr_config = dict(
policy='CosineAnnealing',
min_lr=0.,
warmup='linear',
warmup_iters=10,
warmup_ratio=0.0001, # cannot be 0
warmup_by_epoch=True)
checkpoint_config = dict(interval=10)
# runtime settings
total_epochs = 200

17
docs/MODEL_ZOO.md
View File

@ -2,13 +2,13 @@
## Pre-trained model download links
<table><thead><tr><th>Method</th><th>Config</th><th>Remarks</th><th>Download link</th></tr></thead><tbody><tr><td><a href="https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py" target="_blank" rel="noopener noreferrer">ImageNet</a></td><td>-</td><td>torchvision</td><td><a href="https://drive.google.com/file/d/11xA3TOcbD0qOrwpBfYonEDeseE1wMfBh/view?usp=sharing" target="_blank" rel="noopener noreferrer">imagenet_r50-21352794.pth</a></td></tr><tr><td>Random</td><td>-</td><td>kaiming</td><td><a href="https://drive.google.com/file/d/1UaFTjd6sbKkZEE-f58Zv30bnx7C1qJBb/view?usp=sharing" target="_blank" rel="noopener noreferrer">random_r50-5d0fa71b.pth</a></td></tr><tr><td><a href="https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Doersch_Unsupervised_Visual_Representation_ICCV_2015_paper.pdf" target="_blank" rel="noopener noreferrer">Relative-Loc</a></td><td>selfsup/relative_loc/r50.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1ibk1BI3PFQxZqcxuDfHs3n7JnWKCgl8x/view?usp=sharing" target="_blank" rel="noopener noreferrer">relative_loc_r50-342c9097.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/1803.07728" target="_blank" rel="noopener noreferrer">Rotation-Pred</a></td><td>selfsup/rotation_pred/r50.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1t3oClmIvQ0p8RZ0V5yvQFltzjqBO823Y/view?usp=sharing" target="_blank" rel="noopener noreferrer">rotation_r50-cfab8ebb.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/1807.05520" target="_blank" rel="noopener noreferrer">DeepCluster</a></td><td>selfsup/deepcluster/r50.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1GxgP7pI18JtFxDIC0hnHOanvUYajoLlg/view?usp=sharing" target="_blank" rel="noopener noreferrer">deepcluster_r50-bb8681e2.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/1805.01978" target="_blank" rel="noopener noreferrer">NPID</a></td><td>selfsup/npid/r50.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1sm6I3Y5XnCWdbmeLSF4YupUtPe5nRQMI/view?usp=sharing" target="_blank" rel="noopener noreferrer">npid_r50-dec3df0c.pth</a></td></tr><tr><td><a href="http://openaccess.thecvf.com/content_CVPR_2020/papers/Zhan_Online_Deep_Clustering_for_Unsupervised_Representation_Learning_CVPR_2020_paper.pdf" target="_blank" rel="noopener noreferrer">ODC</a></td><td>selfsup/odc/r50_v1.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1EdhJeZAyMsD_pEW7uMhLzos5xZLdariN/view?usp=sharing" target="_blank" rel="noopener noreferrer">odc_r50_v1-5af5dd0c.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/1911.05722" target="_blank" rel="noopener noreferrer">MoCo</a></td><td>selfsup/moco/r50_v1.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1ANXfnoT8yBQQBBqR_kQLQorK20l65KMy/view?usp=sharing" target="_blank" rel="noopener noreferrer">moco_r50_v1-4ad89b5c.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/2003.04297" target="_blank" rel="noopener noreferrer">MoCo v2</a></td><td>selfsup/moco/r50_v2.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1Cc5qMjPkKW6WeM4ic9Tq-IBxswyJhMnF/view?usp=sharing" target="_blank" rel="noopener noreferrer">moco_r50_v2-58f10cfe.pth</a></td></tr><tr><td></td><td>selfsup/moco/r50_v2_simclr_neck.py</td><td>-&gt; SimCLR neck<br></td><td><a href="https://drive.google.com/file/d/1PnZmCVmFwBv7ZnqMgNYj5DvmbPGM5rCx/view?usp=sharing" target="_blank" rel="noopener noreferrer">moco_r50_v2_simclr_neck-70379356.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/2002.05709" target="_blank" rel="noopener noreferrer">SimCLR</a></td><td>selfsup/simclr/r50_bs256_ep200.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1aZ43nSdivdNxHbM9DKVoZYVhZ8TNnmPp/view?usp=sharing" target="_blank" rel="noopener noreferrer">simclr_r50_bs256_ep200-4577e9a6.pth</a></td></tr><tr><td></td><td>selfsup/simclr/r50_bs256_ep200_mocov2_neck.py</td><td>-&gt; MoCo v2 neck</td><td><a href="https://drive.google.com/file/d/1AXpSKqgWfnj6jCgN65BXSTCKFfuIVELa/view?usp=sharing" target="_blank" rel="noopener noreferrer">simclr_r50_bs256_ep200_mocov2_neck-0d6e5ff2.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/2006.07733" target="_blank" rel="noopener noreferrer">BYOL</a></td><td>selfsup/byol/r50.py</td><td>default</td><td></td></tr></tbody></table>
<table><thead><tr><th>Method</th><th>Config</th><th>Remarks</th><th>Download link</th></tr></thead><tbody><tr><td><a href="https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py" target="_blank" rel="noopener noreferrer">ImageNet</a></td><td>-</td><td>torchvision</td><td><a href="https://drive.google.com/file/d/11xA3TOcbD0qOrwpBfYonEDeseE1wMfBh/view?usp=sharing" target="_blank" rel="noopener noreferrer">imagenet_r50-21352794.pth</a></td></tr><tr><td>Random</td><td>-</td><td>kaiming</td><td><a href="https://drive.google.com/file/d/1UaFTjd6sbKkZEE-f58Zv30bnx7C1qJBb/view?usp=sharing" target="_blank" rel="noopener noreferrer">random_r50-5d0fa71b.pth</a></td></tr><tr><td><a href="https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Doersch_Unsupervised_Visual_Representation_ICCV_2015_paper.pdf" target="_blank" rel="noopener noreferrer">Relative-Loc</a></td><td>selfsup/relative_loc/r50.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1ibk1BI3PFQxZqcxuDfHs3n7JnWKCgl8x/view?usp=sharing" target="_blank" rel="noopener noreferrer">relative_loc_r50-342c9097.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/1803.07728" target="_blank" rel="noopener noreferrer">Rotation-Pred</a></td><td>selfsup/rotation_pred/r50.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1t3oClmIvQ0p8RZ0V5yvQFltzjqBO823Y/view?usp=sharing" target="_blank" rel="noopener noreferrer">rotation_r50-cfab8ebb.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/1807.05520" target="_blank" rel="noopener noreferrer">DeepCluster</a></td><td>selfsup/deepcluster/r50.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1GxgP7pI18JtFxDIC0hnHOanvUYajoLlg/view?usp=sharing" target="_blank" rel="noopener noreferrer">deepcluster_r50-bb8681e2.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/1805.01978" target="_blank" rel="noopener noreferrer">NPID</a></td><td>selfsup/npid/r50.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1sm6I3Y5XnCWdbmeLSF4YupUtPe5nRQMI/view?usp=sharing" target="_blank" rel="noopener noreferrer">npid_r50-dec3df0c.pth</a></td></tr><tr><td><a href="http://openaccess.thecvf.com/content_CVPR_2020/papers/Zhan_Online_Deep_Clustering_for_Unsupervised_Representation_Learning_CVPR_2020_paper.pdf" target="_blank" rel="noopener noreferrer">ODC</a></td><td>selfsup/odc/r50_v1.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1EdhJeZAyMsD_pEW7uMhLzos5xZLdariN/view?usp=sharing" target="_blank" rel="noopener noreferrer">odc_r50_v1-5af5dd0c.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/1911.05722" target="_blank" rel="noopener noreferrer">MoCo</a></td><td>selfsup/moco/r50_v1.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1ANXfnoT8yBQQBBqR_kQLQorK20l65KMy/view?usp=sharing" target="_blank" rel="noopener noreferrer">moco_r50_v1-4ad89b5c.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/2003.04297" target="_blank" rel="noopener noreferrer">MoCo v2</a></td><td>selfsup/moco/r50_v2.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1Cc5qMjPkKW6WeM4ic9Tq-IBxswyJhMnF/view?usp=sharing" target="_blank" rel="noopener noreferrer">moco_r50_v2-58f10cfe.pth</a></td></tr><tr><td></td><td>selfsup/moco/r50_v2_simclr_neck.py</td><td>-&gt; SimCLR neck<br></td><td><a href="https://drive.google.com/file/d/1PnZmCVmFwBv7ZnqMgNYj5DvmbPGM5rCx/view?usp=sharing" target="_blank" rel="noopener noreferrer">moco_r50_v2_simclr_neck-70379356.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/2002.05709" target="_blank" rel="noopener noreferrer">SimCLR</a></td><td>selfsup/simclr/r50_bs256_ep200.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1aZ43nSdivdNxHbM9DKVoZYVhZ8TNnmPp/view?usp=sharing" target="_blank" rel="noopener noreferrer">simclr_r50_bs256_ep200-4577e9a6.pth</a></td></tr><tr><td></td><td>selfsup/simclr/r50_bs256_ep200_mocov2_neck.py</td><td>-&gt; MoCo v2 neck</td><td><a href="https://drive.google.com/file/d/1AXpSKqgWfnj6jCgN65BXSTCKFfuIVELa/view?usp=sharing" target="_blank" rel="noopener noreferrer">simclr_r50_bs256_ep200_mocov2_neck-0d6e5ff2.pth</a></td></tr><tr><td><a href="https://arxiv.org/abs/2006.07733" target="_blank" rel="noopener noreferrer">BYOL</a></td><td>selfsup/byol/r50_bs4096_ep200.py</td><td>default</td><td><a href="https://drive.google.com/file/d/1Whj3j5E3ShQj_VufjrJSzWiq1xcZZCXN/view?usp=sharing" target="_blank" rel="noopener noreferrer">byol_r50-e3b0c442.pth</a></td></tr></tbody></table>
## Benchmarks
### VOC07 SVM & SVM Low-shot
<table><thead><tr><th rowspan="2">Method</th><th rowspan="2">Config</th><th rowspan="2">Remarks</th><th rowspan="2">Best layer</th><th rowspan="2">VOC07 SVM</th><th colspan="8">VOC07 SVM Low-shot</th></tr><tr><td>1</td><td>2</td><td>4</td><td>8</td><td>16</td><td>32</td><td>64</td><td>96</td></tr></thead><tbody><tr><td><a href="https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py" target="_blank" rel="noopener noreferrer">ImageNet</a></td><td>-</td><td>torchvision</td><td>feat5</td><td>87.17</td><td>52.99</td><td>63.55</td><td>73.7</td><td>78.79</td><td>81.76</td><td>83.75</td><td>85.18</td><td>85.97</td></tr><tr><td>Random</td><td>-</td><td>kaiming</td><td>feat2</td><td>30.22</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Doersch_Unsupervised_Visual_Representation_ICCV_2015_paper.pdf" target="_blank" rel="noopener noreferrer">Relative-Loc</a></td><td></td><td>default</td><td>feat4</td><td>65.06</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1803.07728" target="_blank" rel="noopener noreferrer">Rotation-Pred</a></td><td>selfsup/rotation_pred/r50.py</td><td>default</td><td>feat4</td><td>67.38</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1807.05520" target="_blank" rel="noopener noreferrer">DeepCluster</a></td><td>selfsup/deepcluster/r50.py</td><td>default</td><td>feat5</td><td>74.26</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1805.01978" target="_blank" rel="noopener noreferrer">NPID</a></td><td>selfsup/npid/r50.py</td><td>default</td><td>feat5</td><td>74.50</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td></td><td>selfsup/npid/r50_ensure_neg.py</td><td>ensure_neg=True</td><td>feat5</td><td>75.70</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="http://openaccess.thecvf.com/content_CVPR_2020/papers/Zhan_Online_Deep_Clustering_for_Unsupervised_Representation_Learning_CVPR_2020_paper.pdf" target="_blank" rel="noopener noreferrer">ODC</a></td><td>selfsup/odc/r50_v1.py</td><td>default</td><td>feat5</td><td>78.42</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1911.05722" target="_blank" rel="noopener noreferrer">MoCo</a></td><td>selfsup/moco/r50_v1.py</td><td>default</td><td>feat5</td><td>79.18</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/2003.04297" target="_blank" rel="noopener noreferrer">MoCo v2</a></td><td>selfsup/moco/r50_v2.py</td><td>default</td><td>feat5</td><td>84.05</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td></td><td>selfsup/moco/r50_v2_simclr_neck.py</td><td>-&gt; SimCLR neck<br></td><td>feat5</td><td>84.00</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/2002.05709" target="_blank" rel="noopener noreferrer">SimCLR</a></td><td>selfsup/simclr/r50_bs256_ep200.py</td><td>default</td><td>feat5</td><td>78.95</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td></td><td>selfsup/simclr/r50_bs256_ep200_mocov2_neck.py</td><td>-&gt; MoCo v2 neck</td><td>feat5</td><td>77.65</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/2006.07733" target="_blank" rel="noopener noreferrer">BYOL</a></td><td>selfsup/byol/r50.py</td><td>default</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr></tbody></table>
<table><thead><tr><th rowspan="2">Method</th><th rowspan="2">Config</th><th rowspan="2">Remarks</th><th rowspan="2">Best layer</th><th rowspan="2">VOC07 SVM</th><th colspan="8">VOC07 SVM Low-shot</th></tr><tr><td>1</td><td>2</td><td>4</td><td>8</td><td>16</td><td>32</td><td>64</td><td>96</td></tr></thead><tbody><tr><td><a href="https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py" target="_blank" rel="noopener noreferrer">ImageNet</a></td><td>-</td><td>torchvision</td><td>feat5</td><td>87.17</td><td>52.99</td><td>63.55</td><td>73.7</td><td>78.79</td><td>81.76</td><td>83.75</td><td>85.18</td><td>85.97</td></tr><tr><td>Random</td><td>-</td><td>kaiming</td><td>feat2</td><td>30.22</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Doersch_Unsupervised_Visual_Representation_ICCV_2015_paper.pdf" target="_blank" rel="noopener noreferrer">Relative-Loc</a></td><td></td><td>default</td><td>feat4</td><td>65.06</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1803.07728" target="_blank" rel="noopener noreferrer">Rotation-Pred</a></td><td>selfsup/rotation_pred/r50.py</td><td>default</td><td>feat4</td><td>67.38</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1807.05520" target="_blank" rel="noopener noreferrer">DeepCluster</a></td><td>selfsup/deepcluster/r50.py</td><td>default</td><td>feat5</td><td>74.26</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1805.01978" target="_blank" rel="noopener noreferrer">NPID</a></td><td>selfsup/npid/r50.py</td><td>default</td><td>feat5</td><td>74.50</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td></td><td>selfsup/npid/r50_ensure_neg.py</td><td>ensure_neg=True</td><td>feat5</td><td>75.70</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="http://openaccess.thecvf.com/content_CVPR_2020/papers/Zhan_Online_Deep_Clustering_for_Unsupervised_Representation_Learning_CVPR_2020_paper.pdf" target="_blank" rel="noopener noreferrer">ODC</a></td><td>selfsup/odc/r50_v1.py</td><td>default</td><td>feat5</td><td>78.42</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1911.05722" target="_blank" rel="noopener noreferrer">MoCo</a></td><td>selfsup/moco/r50_v1.py</td><td>default</td><td>feat5</td><td>79.18</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/2003.04297" target="_blank" rel="noopener noreferrer">MoCo v2</a></td><td>selfsup/moco/r50_v2.py</td><td>default</td><td>feat5</td><td>84.05</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td></td><td>selfsup/moco/r50_v2_simclr_neck.py</td><td>-&gt; SimCLR neck<br></td><td>feat5</td><td>84.00</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/2002.05709" target="_blank" rel="noopener noreferrer">SimCLR</a></td><td>selfsup/simclr/r50_bs256_ep200.py</td><td>default</td><td>feat5</td><td>78.95</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td></td><td>selfsup/simclr/r50_bs256_ep200_mocov2_neck.py</td><td>-&gt; MoCo v2 neck</td><td>feat5</td><td>77.65</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/2006.07733" target="_blank" rel="noopener noreferrer">BYOL</a></td><td>selfsup/byol/r50_bs4096_ep200.py</td><td>default</td><td>feat5</td><td>85.10</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr></tbody></table>
### ImageNet Linear Classification
@ -18,7 +18,7 @@
* ImageNet (Multi) evaluates features in around 9k dimensions from different layers. Top-1 result of the last epoch is reported.
* ImageNet (Last) evaluates the last feature after global average pooling, e.g., 2048 dimensions for resnet50. The best top-1 result among all epochs is reported.
<table><thead><tr><th rowspan="2">Method</th><th rowspan="2">Config</th><th rowspan="2">Remarks</th><th colspan="5">ImageNet (Multi)</th><th>ImageNet (Last)</th></tr><tr><td>feat1</td><td>feat2</td><td>feat3</td><td>feat4</td><td>feat5</td><td>avgpool</td></tr></thead><tbody><tr><td><a href="https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py" target="_blank" rel="noopener noreferrer">ImageNet</a></td><td>-</td><td>torchvision</td><td>15.18</td><td>33.96</td><td>47.86</td><td>67.56</td><td>76.17</td><td>74.12</td></tr><tr><td>Random</td><td>-</td><td>kaiming</td><td>11.37</td><td>16.21</td><td>13.47</td><td>9.07</td><td>6.54</td><td>4.35</td></tr><tr><td><a href="https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Doersch_Unsupervised_Visual_Representation_ICCV_2015_paper.pdf" target="_blank" rel="noopener noreferrer">Relative-Loc</a></td><td>selfsup/relative_loc/r50.py</td><td>default</td><td>14.76</td><td>31.29</td><td>45.77</td><td>49.31</td><td>40.20</td><td>38.83</td></tr><tr><td><a href="https://arxiv.org/abs/1803.07728" target="_blank" rel="noopener noreferrer">Rotation-Pred</a></td><td>selfsup/rotation_pred/r50.py</td><td>default</td><td>12.89</td><td>34.30</td><td>44.91</td><td>54.99</td><td>49.09</td><td>47.01</td></tr><tr><td><a href="https://arxiv.org/abs/1807.05520" target="_blank" rel="noopener noreferrer">DeepCluster</a></td><td>selfsup/deepcluster/r50.py</td><td>default</td><td>12.78</td><td>30.81</td><td>43.88</td><td>57.71</td><td>51.68</td><td>46.92</td></tr><tr><td><a href="https://arxiv.org/abs/1805.01978" target="_blank" rel="noopener noreferrer">NPID</a></td><td>selfsup/npid/r50.py</td><td>default</td><td>14.28</td><td>31.20</td><td>40.68</td><td>54.46</td><td>56.61</td><td>56.60</td></tr><tr><td><a href="http://openaccess.thecvf.com/content_CVPR_2020/papers/Zhan_Online_Deep_Clustering_for_Unsupervised_Representation_Learning_CVPR_2020_paper.pdf" target="_blank" rel="noopener noreferrer">ODC</a></td><td>selfsup/odc/r50_v1.py</td><td>default</td><td>14.76</td><td>31.82</td><td>42.44</td><td>55.76</td><td>57.70</td><td>53.42</td></tr><tr><td><a href="https://arxiv.org/abs/1911.05722" target="_blank" rel="noopener noreferrer">MoCo</a></td><td>selfsup/moco/r50_v1.py</td><td>default</td><td>15.32</td><td>33.08</td><td>44.68</td><td>57.27</td><td>60.60</td><td>61.02</td></tr><tr><td><a href="https://arxiv.org/abs/2003.04297" target="_blank" rel="noopener noreferrer">MoCo v2</a></td><td>selfsup/moco/r50_v2.py</td><td>default</td><td>15.35</td><td>34.57</td><td>45.81</td><td>60.96</td><td>66.72</td><td>67.02</td></tr><tr><td></td><td>selfsup/moco/r50_v2_simclr_neck.py</td><td>-&gt; SimCLR neck<br></td><td>15.19</td><td>32.54</td><td>43.12</td><td>60.36</td><td>67.08</td><td>65.39</td></tr><tr><td><a href="https://arxiv.org/abs/2002.05709" target="_blank" rel="noopener noreferrer">SimCLR</a></td><td>selfsup/simclr/r50_bs256_ep200.py</td><td>default</td><td>17.09</td><td>31.37</td><td>41.38</td><td>54.35</td><td>61.57</td><td>60.06</td></tr><tr><td></td><td>selfsup/simclr/r50_bs256_ep200_mocov2_neck.py</td><td>-&gt; MoCo v2 neck</td><td>16.97</td><td>31.88</td><td>41.73</td><td>54.33</td><td>59.94</td><td>58.00</td></tr><tr><td><a href="https://arxiv.org/abs/2006.07733" target="_blank" rel="noopener noreferrer">BYOL</a></td><td>selfsup/byol/r50.py</td><td>default</td><td></td><td></td><td></td><td></td><td></td><td></td></tr></tbody></table>
<table><thead><tr><th rowspan="2">Method</th><th rowspan="2">Config</th><th rowspan="2">Remarks</th><th colspan="5">ImageNet (Multi)</th><th>ImageNet (Last)</th></tr><tr><td>feat1</td><td>feat2</td><td>feat3</td><td>feat4</td><td>feat5</td><td>avgpool</td></tr></thead><tbody><tr><td><a href="https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py" target="_blank" rel="noopener noreferrer">ImageNet</a></td><td>-</td><td>torchvision</td><td>15.18</td><td>33.96</td><td>47.86</td><td>67.56</td><td>76.17</td><td>74.12</td></tr><tr><td>Random</td><td>-</td><td>kaiming</td><td>11.37</td><td>16.21</td><td>13.47</td><td>9.07</td><td>6.54</td><td>4.35</td></tr><tr><td><a href="https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Doersch_Unsupervised_Visual_Representation_ICCV_2015_paper.pdf" target="_blank" rel="noopener noreferrer">Relative-Loc</a></td><td>selfsup/relative_loc/r50.py</td><td>default</td><td>14.76</td><td>31.29</td><td>45.77</td><td>49.31</td><td>40.20</td><td>38.83</td></tr><tr><td><a href="https://arxiv.org/abs/1803.07728" target="_blank" rel="noopener noreferrer">Rotation-Pred</a></td><td>selfsup/rotation_pred/r50.py</td><td>default</td><td>12.89</td><td>34.30</td><td>44.91</td><td>54.99</td><td>49.09</td><td>47.01</td></tr><tr><td><a href="https://arxiv.org/abs/1807.05520" target="_blank" rel="noopener noreferrer">DeepCluster</a></td><td>selfsup/deepcluster/r50.py</td><td>default</td><td>12.78</td><td>30.81</td><td>43.88</td><td>57.71</td><td>51.68</td><td>46.92</td></tr><tr><td><a href="https://arxiv.org/abs/1805.01978" target="_blank" rel="noopener noreferrer">NPID</a></td><td>selfsup/npid/r50.py</td><td>default</td><td>14.28</td><td>31.20</td><td>40.68</td><td>54.46</td><td>56.61</td><td>56.60</td></tr><tr><td><a href="http://openaccess.thecvf.com/content_CVPR_2020/papers/Zhan_Online_Deep_Clustering_for_Unsupervised_Representation_Learning_CVPR_2020_paper.pdf" target="_blank" rel="noopener noreferrer">ODC</a></td><td>selfsup/odc/r50_v1.py</td><td>default</td><td>14.76</td><td>31.82</td><td>42.44</td><td>55.76</td><td>57.70</td><td>53.42</td></tr><tr><td><a href="https://arxiv.org/abs/1911.05722" target="_blank" rel="noopener noreferrer">MoCo</a></td><td>selfsup/moco/r50_v1.py</td><td>default</td><td>15.32</td><td>33.08</td><td>44.68</td><td>57.27</td><td>60.60</td><td>61.02</td></tr><tr><td><a href="https://arxiv.org/abs/2003.04297" target="_blank" rel="noopener noreferrer">MoCo v2</a></td><td>selfsup/moco/r50_v2.py</td><td>default</td><td>15.35</td><td>34.57</td><td>45.81</td><td>60.96</td><td>66.72</td><td>67.02</td></tr><tr><td></td><td>selfsup/moco/r50_v2_simclr_neck.py</td><td>-&gt; SimCLR neck<br></td><td>15.19</td><td>32.54</td><td>43.12</td><td>60.36</td><td>67.08</td><td>65.39</td></tr><tr><td><a href="https://arxiv.org/abs/2002.05709" target="_blank" rel="noopener noreferrer">SimCLR</a></td><td>selfsup/simclr/r50_bs256_ep200.py</td><td>default</td><td>17.09</td><td>31.37</td><td>41.38</td><td>54.35</td><td>61.57</td><td>60.06</td></tr><tr><td></td><td>selfsup/simclr/r50_bs256_ep200_mocov2_neck.py</td><td>-&gt; MoCo v2 neck</td><td>16.97</td><td>31.88</td><td>41.73</td><td>54.33</td><td>59.94</td><td>58.00</td></tr><tr><td><a href="https://arxiv.org/abs/2006.07733" target="_blank" rel="noopener noreferrer">BYOL</a></td><td>selfsup/byol/r50_bs4096_ep200.py</td><td>default</td><td></td><td></td><td></td><td></td><td></td><td>67.10</td></tr></tbody></table>
### Places205 Linear Classification
@ -44,3 +44,14 @@ Coming soon.
* Please follow [here](GETTING_STARTED.md#voc0712--coco17-object-detection) to run the evaluation.
<table><thead><tr><th rowspan="2">Method</th><th rowspan="2">Config</th><th rowspan="2">Remarks</th><th colspan="3">VOC07+12</th></tr><tr><td>AP50</td><td>AP</td><td>AP75</td></tr></thead><tbody><tr><td><a href="https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py" target="_blank" rel="noopener noreferrer">ImageNet</a></td><td>-</td><td>torchvision</td><td>81.58</td><td>54.19</td><td>59.80</td></tr><tr><td>Random</td><td>-</td><td>kaiming</td><td>59.02</td><td>32.83</td><td>31.60</td></tr><tr><td><a href="https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Doersch_Unsupervised_Visual_Representation_ICCV_2015_paper.pdf" target="_blank" rel="noopener noreferrer">Relative-Loc</a></td><td>selfsup/relative_loc/r50.py</td><td>default</td><td></td><td></td><td></td></tr><tr><td><a href="https://arxiv.org/abs/1803.07728" target="_blank" rel="noopener noreferrer">Rotation-Pred</a></td><td>selfsup/rotation_pred/r50.py</td><td>default</td><td>80.91</td><td>55.52</td><td>61.39</td></tr><tr><td><a href="https://arxiv.org/abs/1805.01978" target="_blank" rel="noopener noreferrer">NPID</a></td><td>selfsup/npid/r50.py</td><td>default</td><td>80.03</td><td>54.11</td><td>59.50</td></tr><tr><td><a href="https://arxiv.org/abs/1911.05722" target="_blank" rel="noopener noreferrer">MoCo</a></td><td>selfsup/moco/r50_v1.py</td><td>default</td><td>81.38</td><td>55.95</td><td>62.23</td></tr><tr><td><a href="https://arxiv.org/abs/2003.04297" target="_blank" rel="noopener noreferrer">MoCo v2</a></td><td>selfsup/moco/r50_v2.py</td><td>default</td><td>81.96</td><td>56.63</td><td>62.90</td></tr><tr><td><a href="https://arxiv.org/abs/2002.05709" target="_blank" rel="noopener noreferrer">SimCLR</a></td><td>selfsup/simclr/r50_bs256_ep200.py</td><td>default</td><td>79.41</td><td>51.54</td><td>55.63</td></tr><tr><td><a href="https://arxiv.org/abs/2006.07733" target="_blank" rel="noopener noreferrer">BYOL</a></td><td>selfsup/byol/r50.py</td><td>default</td><td></td><td></td><td></td></tr></tbody></table>
### COCO2017 Object Detection
**Note**
* This benchmark follows the evluation protocols set up by MoCo.
* Config: `benchmarks/detection/configs/coco_R_50_C4_2x_moco.yaml`.
* Please follow [here](GETTING_STARTED.md#voc0712--coco17-object-detection) to run the evaluation.
<table><thead><tr><th rowspan="2">Method</th><th rowspan="2">Config</th><th rowspan="2">Remarks</th><th colspan="6">COCO2017</th></tr><tr><td>AP50(Box)</td><td>AP(Box)</td><td>AP75(Box)</td><td>AP50(Mask)</td><td>AP(Mask)</td><td>AP75(Mask)</td></tr></thead><tbody><tr><td><a href="https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py" target="_blank" rel="noopener noreferrer">ImageNet</a></td><td>-</td><td>torchvision</td><td>59.9</td><td>40.0</td><td>43.1</td><td>56.5</td><td>34.7</td><td>36.9</td></tr><tr><td>Random</td><td>-</td><td>kaiming</td><td>54.6</td><td>35.6</td><td>38.2</td><td>51.5</td><td>31.4</td><td>33.5</td></tr><tr><td><a href="https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Doersch_Unsupervised_Visual_Representation_ICCV_2015_paper.pdf" target="_blank" rel="noopener noreferrer">Relative-Loc</a></td><td>selfsup/relative_loc/r50.py</td><td>default</td><td>59.6</td><td>40.0</td><td>43.5</td><td>56.5</td><td>35.0</td><td>37.3</td></tr><tr><td><a href="https://arxiv.org/abs/1803.07728" target="_blank" rel="noopener noreferrer">Rotation-Pred</a></td><td>selfsup/rotation_pred/r50.py</td><td>default</td><td>59.3</td><td>40.0</td><td>43.6</td><td>56.0</td><td>34.9</td><td>37.4</td></tr><tr><td><a href="https://arxiv.org/abs/1805.01978" target="_blank" rel="noopener noreferrer">NPID</a></td><td>selfsup/npid/r50.py</td><td>default</td><td>59.0</td><td>39.4</td><td>42.8</td><td>55.9</td><td>34.5</td><td>36.6</td></tr><tr><td><a href="https://arxiv.org/abs/1911.05722" target="_blank" rel="noopener noreferrer">MoCo</a></td><td>selfsup/moco/r50_v1.py</td><td>default</td><td>60.5</td><td>40.9</td><td>44.2</td><td>57.1</td><td>35.5</td><td>37.7</td></tr><tr><td><a href="https://arxiv.org/abs/2003.04297">MoCo v2</a></td><td>selfsup/moco/r50_v2.py</td><td>default</td><td>60.7</td><td>40.9</td><td>44.2</td><td>57.2</td><td>35.5</td><td>37.9</td></tr><tr><td><a href="https://arxiv.org/abs/2002.05709">SimCLR</a></td><td>selfsup/simclr/r50_bs256_ep200.py</td><td>default</td><td>59.1</td><td>39.6</td><td>42.9</td><td>55.9</td><td>34.6</td><td>37.1</td></tr><tr><td><a href="https://arxiv.org/abs/2006.07733">BYOL</a></td><td>selfsup/byol/r50.py</td><td>default</td><td></td><td></td><td></td><td></td><td></td><td></td></tr></tbody></table>