SimCLR

Implementation of SimCLR (A Simple Framework for Contrastive Learning of Visual Representations)

• Requirements

  • numpy
  • torch
  • torchvision
  • opencv-python

• command

 - python3 main.py --epochs [epochs] --batch_size [B] --temperature [T] --strength [S] --out_dim [out_dim] --num_worker [N] --valid_size [val_size] - python3 linear_eval.py --batch_size [B] --simclr_path [path] --dataset [dataset] --hid_dim [hid_dim] --num_worker [N] --finetune --baseline

• results

Dataset	STL10	CIFAR10
Baseline	54.801	70.653
No Finetune	66.192	50.401
Finetune	73.866	71.795

Reference

1. A Simple Framework for Contrastive Learning of Visual Representations (https://arxiv.org/abs/2002.05709)
2. STL-10 Dataset (https://cs.stanford.edu/~acoates/stl10/)