DisCo: Remedy Self-supervised Learning on Lightweight Models with Distilled Contrastive Learning
Pytorch Implementation for DisCo: Remedy Self-supervised Learning on Lightweight Models with Distilled Contrastive Learning
If the project is useful to you, please give us a star.
⭐️
@article{gao2021disco,
title={DisCo: Remedy Self-supervised Learning on Lightweight Models with Distilled Contrastive Learning},
author={Gao, Yuting and Zhuang, Jia-Xin and Li, Ke and Cheng, Hao and Guo, Xiaowei and Huang, Feiyue and Ji, Rongrong and Sun, Xing},
journal={arXiv preprint arXiv:2104.09124},
year={2021}
}
Checkpoints
Teacher Models
| Architecture | Self-supervised Methods | Model Checkpoints |
|---|---|---|
| ResNet152 | MoCo-V2 | Model |
| ResNet101 | MoCo-V2 | Model |
| ResNet50 | MoCo-V2 | Model |
For teacher models such as ResNet-50*2 etc, we use their official implementation, which can be downloaded from their github pages.
Student Models by DisCo
| Teacher/Students | Efficient-B0 | ResNet-18 | Vit-Tiny | XCiT-Tiny |
|---|---|---|---|---|
| ResNet-50 | Model | Model | - | - |
| ResNet-101 | Model | Model | - | - |
| ResNet-152 | Model | Model | - | - |
| ResNet-50*2 | Model | Model | - | - |
| ViT-Small | - | - | Model | - |
| XCiT-Small | - | - | - | Model |
Requirements
-
Python3
-
Pytorch 1.6+
-
Detectron2
-
8 GPUs are preferred
-
ImageNet, Cifar10/100, VOC, COCO
Run
Before running, we firstly move all data into share memory
cp /path/to/ImageNet /dev/shm
Pretrain Model
For pretraining baseline models with default hidden layer dimension in Tab1
# Switch to moco directory
cd moco
# R-50
python3 -u main_moco.py -a resnet50 --batch-size 256 --learning-rate 0.03 --mlp --moco-t 0.2 --aug-plus --cos --epochs 200 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --hidden 2048 /dev/shm/ 2>&1 | tee ./logs/std.log
python3 main_lincls.py -a resnet50 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
# R-101
python3 -u main_moco.py -a resnet101 --batch-size 256 --learning-rate 0.03 --mlp --moco-t 0.2 --aug-plus --cos --epochs 200 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --hidden 2048 /dev/shm/ 2>&1 | tee ./logs/std.log
python3 main_lincls.py -a resnet101 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
# R-152
python3 -u main_moco.py -a resnet152 --batch-size 256 --learning-rate 0.03 --mlp --moco-t 0.2 --aug-plus --cos --epochs 800 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --hidden 2048 /dev/shm/ 2>&1 | tee ./logs/std.log
python3 main_lincls.py -a resnet152 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0799.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
# Mob
python3 -u main_moco.py -a mobilenetv3 --batch-size 256 --learning-rate 0.03 --mlp --moco-t 0.2 --aug-plus --cos --epochs 200 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --hidden 512 /dev/shm 2>&1 | tee ./logs/std.log
# Evaluation
python3 main_lincls.py -a mobilenetv3 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
# Effi-B0
python3 -u main_moco.py -a efficientb0 --batch-size 256 --learning-rate 0.03 --mlp --moco-t 0.2 --aug-plus --cos --epochs 200 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --hidden 1280 2>&1 | tee ./logs/std.log
# Evaluation
python3 main_lincls.py -a efficientb0 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
# Effi-B1
python3 -u main_moco.py -a efficientb1 --batch-size 256 --learning-rate 0.03 --mlp --moco-t 0.2 --aug-plus --cos --epochs 200 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --hidden 1280 /dev/shm 2>&1 | tee ./logs/std.log
# Evaluation
python3 main_lincls.py -a efficientb1 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
# R-18
python3 -u main_moco.py -a resnet18 --batch-size 256 --learning-rate 0.03 --mlp --moco-t 0.2 --aug-plus --cos --epochs 200 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --hidden 1280 /dev/shm/ 2>&1 | tee ./logs/std.log
# Evaluation
python3 main_lincls.py -a resnet18 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
# R-34
python3 -u main_moco.py -a resnet34 --batch-size 256 --learning-rate 0.03 --mlp --moco-t 0.2 --aug-plus --cos --epochs 200 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --hidden 1280 /dev/shm/ 2>&1 | tee ./logs/std.log
# Evaluation
python3 main_lincls.py -a resnet34 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
DisCo
For training DisCo in Tab1, Comparision with baseline
# Switch to DisCo directory
cd DisCo
# R-50 & Effib0
python3 -u main.py -a efficientb0 --lr 0.03 --batch-size 256 --moco-t 0.2 --aug-plus --dist-url 'tcp://localhost:10043' --multiprocessing-distributed --world-size 1 --rank 0 --mlp --cos --teacher_arch resnet50 --teacher /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm/ 2>&1 | tee ./logs/std.log
# Evaluation
python3 -u main_lincls.py -a efficientb0 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
# R50w2 & Effib0
python3 -u main.py -a efficientb0 --lr 0.03 --batch-size 256 --moco-t 0.2 --aug-plus --dist-url 'tcp://localhost:10043' --multiprocessing-distributed --world-size 1 --rank 0 --mlp --cos --teacher_arch resnet50w2 --teacher /path/to/swav_RN50w2_400ep_pretrain.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
# Evaluation
python3 yt_main_lincls.py -a resnet18 --learning-rate 30.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
For Tab2, Linear evaluation top-1 accuracy (%) on ImageNet compared with different distillation methods.
# RKD+DisCo, Eff-b0
python3 -u main_moco_distill_rkd.py -a efficientb0 --lr 0.03 --batch-size 256 --moco-t 0.2 --aug-plus --dist-url 'tcp://localhost:10043' --multiprocessing-distributed --world-size 1 --rank 0 --mlp --cos --teacher /path/to/teacher_res50.pth.tar --use-mse /dev/shm 2>&1 | tee ./logs/std.log
# Evaluation
python3 -u main_lincls.py -a efficientb0 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
# RKD, Eff-b0
python3 -u main_moco_distill_rkd.py -a efficientb0 --lr 0.03 --batch-size 256 --moco-t 0.2 --aug-plus --dist-url 'tcp://localhost:10043' --multiprocessing-distributed --world-size 1 --rank 0 --mlp --cos --teacher /path/to/teacher_res50.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
# Evaluation
python3 -u main_lincls.py -a efficientb0 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
For Tab3 , **Object detection and instance segmentation results **
# Cp data to /dev/shm and set up path for Detectron2
cp -r /path/to/VOCdevkit/* /dev/shm/
cp -r /path/to/coco_2017 /dev/shm/coco
export DETECTRON2_DATASETS=/dev/shm
pip install /youtu-reid/jiaxzhuang/acmm/detectron2-0.4+cu101-cp36-cp36m-linux_x86_64.whl
cd detection
# Convert model for Detectron2
python3 convert-pretrain-to-detectron2.py /path/ckpt/checkpoint_0199.pth.tar ./output.pkl
# Evaluation on VOC
python3 train_net.py --config-file configs/pascal_voc_R_50_C4_24k_moco.yaml --num-gpus 8 --resume MODEL.RESNETS.DEPTH 34 MODEL.RESNETS.RES2_OUT_CHANNELS 64 2>&1 | tee ../logs/std.log
# Evaluation on CoCo
python3 train_net.py --config-file configs/coco_R_50_C4_2x_moco.yaml --num-gpus 8 --resume MODEL.RESNETS.DEPTH 18 MODEL.RESNETS.RES2_OUT_CHANNELS 64 2>&1 | tee ../logs/std.log
For Fig5 , evaluation on Semi-Supervised Tasks
# Copy 1%, 10% ImageNet from the complete ImageNet, according to split from SimCLR.
cd data
# Need to set up path to Compelete ImageNet and the output path.
python3 -u imagenet_1_fraction.py --ratio 1
python3 -u imagenet_1_fraction.py --ratio 10
# Evaluation on 1% ImageNet with Eff-B0 by DisCo
cp -r /path/to/imagenet_1_fraction/train /dev/shm
cp -r /path/to/imagenet_1_fraction/val /dev/shm/
python3 -u main_lincls_semi.py -a efficientb0 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
# Evaluation on 10% ImageNet with R-18 by DisCo
cp -r /path/to/imagenet_10_fraction/train /dev/shm
cp -r /path/to/imagenet_10_fraction/val /dev/shm/
python3 -u main_lincls_semi.py -a resnet18 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
For Fig6, evaluation on Cifar10/Cifar100
# Copy Cifar10/100 to /dev/shm
cp /path/to/Cifar10/100 /dev/shm
# Evaluation on 1% Cifar10 with Eff-B0 by DisCo
python3 cifar_main_lincls.py -a efficientb0 --dataset cifar10 --lr 3 --epochs 200 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
# Evaluation on Cifar100 with Resnet18 by DisCo
python3 cifar_main_lincls.py -a resnet18 --dataset cifar100 --lr 3 --epochs 200 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
For Tab4, Linear evaluation top-1 accuracy (%) on ImageNet, compared with SEED with consistent dimension in hidden layer.
python3 -u main.py -a efficientb0 --lr 0.03 --batch-size 256 --moco-t 0.2 --aug-plus --dist-url 'tcp://localhost:10043' --multiprocessing-distributed --world-size 1 --rank 0 --mlp --cos --teacher_arch resnet50 --teacher /path/to/ckpt/checkpoint_0199.pth.tar --hidden 2048 /dev/shm/ 2>&1 | tee ./logs/std.log
# Evaluation
python3 -u main_lincls.py -a efficientb0 --learning-rate 3.0 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --pretrained /path/to/ckpt/checkpoint_0199.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
For Tab5, Linear evaluation top-1 accuracy (%) on ImageNet with SwAV as the testbed.
# SwAV, Train with SwAV only
cd swav-master
python3 -m torch.distributed.launch --nproc_per_node=8 main_swav.py \
--data_path /dev/shm/train \
--base_lr 0.6 \
--final_lr 0.0006 \
--warmup_epochs 0 \
--crops_for_assign 0 1 \
--size_crops 224 96 \
--nmb_crops 2 6 \
--min_scale_crops 0.14 0.05 \
--max_scale_crops 1. 0.14 \
--use_fp16 true \
--freeze_prototypes_niters 5005 \
--queue_length 3840 \
--epoch_queue_starts 15 \
--dump_path ./ckpt \
--sync_bn pytorch \
--temperature 0.1 \
--epsilon 0.05 \
--sinkhorn_iterations 3 \
--feat_dim 128 \
--nmb_prototypes 3000 \
--epochs 200 \
--batch_size 64 \
--wd 0.000001 \
--arch efficientb0 \
--use_fp16 true 2>&1 | tee ./logs/std.log
# Evaluation
python3 -m torch.distributed.launch --nproc_per_node=8 eval_linear.py --arch efficientb0 --data_path /dev/shm --pretrained /path/to/ckpt/checkpoints/ckp-199.pth 2>&1 | tee ./logs/std.log
# DisCo + SwAV
python3 -m torch.distributed.launch --nproc_per_node=8 main_swav_distill.py \
--data_path /dev/shm/train \
--base_lr 0.6 \
--final_lr 0.0006 \
--warmup_epochs 0 \
--crops_for_assign 0 1 \
--size_crops 224 96 \
--nmb_crops 2 6 \
--min_scale_crops 0.14 0.05 \
--max_scale_crops 1. 0.14 \
--use_fp16 true \
--freeze_prototypes_niters 5005 \
--queue_length 3840 \
--epoch_queue_starts 15 \
--dump_path ./ckpt \
--sync_bn pytorch \
--temperature 0.1 \
--epsilon 0.05 \
--sinkhorn_iterations 3 \
--feat_dim 128 \
--nmb_prototypes 3000 \
--epochs 200 \
--batch_size 64 \
--wd 0.000001 \
--arch efficientb0 \
--pretrained /path/to/swav_800ep_pretrain.pth.tar 2>&1 | tee ./logs/std.log
For Tab6, Linear evaluation top-1 accuracy (%) on ImageNet with variants of teacher pre-training methods.
# SwAV
python3 -u main.py -a resnet34 --lr 0.03 --batch-size 256 --moco-t 0.2 --aug-plus --dist-url 'tcp://localhost:10043' --multiprocessing-distributed --world-size 1 --rank 0 --mlp --cos --teacher_arch SWAVresnet50 --teacher /path/to/swav_800ep_pretrain.pth.tar /dev/shm 2>&1 | tee ./logs/std.log
Visualization
cd DisCo
# Generate Embed
# Move Embed to data path
python -u draw.py
Thanks
Code heavily depends on MoCo-V2, Detectron2.
44 Dec 17, 2022
49 Dec 21, 2022
42 Nov 24, 2022
64 Dec 16, 2022
282 Jan 9, 2023
157 Dec 16, 2022
107 Dec 14, 2022
342 Jan 3, 2023
92 Nov 30, 2022
18 Dec 17, 2022
28 Dec 12, 2022
478 Dec 25, 2022
491 Jan 3, 2023
59 Dec 28, 2022
4 Jul 12, 2021
27 Dec 10, 2022
32 Jan 2, 2023
13 Nov 30, 2022