SimCLR

A PyTorch implementation of SimCLR based on ICML 2020 paper A Simple Framework for Contrastive Learning of Visual Representations.

Requirements

• Anaconda
• PyTorch

conda install pytorch torchvision cudatoolkit=10.0 -c pytorch

• thop

pip install thop

Dataset

CIFAR10 dataset is used in this repo, the dataset will be downloaded into data directory by PyTorch automatically.

Usage

Train SimCLR

python main.py --batch_size 1024 --epochs 1000 
optional arguments:
--feature_dim                 Feature dim for latent vector [default value is 128]
--temperature                 Temperature used in softmax [default value is 0.5]
--k                           Top k most similar images used to predict the label [default value is 200]
--batch_size                  Number of images in each mini-batch [default value is 512]
--epochs                      Number of sweeps over the dataset to train [default value is 500]

Linear Evaluation

python linear.py --batch_size 1024 --epochs 200 
optional arguments:
--model_path                  The pretrained model path [default value is 'results/128_0.5_200_512_500_model.pth']
--batch_size                  Number of images in each mini-batch [default value is 512]
--epochs                      Number of sweeps over the dataset to train [default value is 100]

Results

There are some difference between this implementation and official implementation, the model (ResNet50) is trained on one NVIDIA TESLA V100(32G) GPU:

1. No Gaussian blur used;
2. Adam optimizer with learning rate 1e-3 is used to replace LARS optimizer;
3. No Linear learning rate scaling used;
4. No Linear Warmup and CosineLR Schedule used.

Evaluation Protocol

Params (M)

FLOPs (G)

Feature Dim

Batch Size

Epoch Num

τ

K

Top1 Acc %

Top5 Acc %

Download

KNN

24.62

1.31

128

512

500

0.5

200

89.1

99.6

model | gc5k

Linear

23.52

1.30

512

100

92.0

99.8

model | f7j2