MoCo v2 Implementation with Pytorch

• Unofficial implementation of the paper Improved Baselines with Momentum Contrastive Learning

0. Develop Environment

• Check env.txt for more details

1. Implementation Details

• data.py : data augmentations, dataset
• main_linear.py : only train linear classifier with frozen backbone
• main_pretrain.py : pre-train backbone
• models_linear.py : model for linear probing
• models_moco.py : model for pre-train
• utils.py : utils such as scheduler, logger
• logs : log files
• visualize : visualize pretrain, linear probing log (acc, lr, train loss, val loss)

2. Linear Probing Result Comparison on ImageNet

Source	Score	Detail
Paper	67.7	200 epochs
Paper	71.1	800 epochs
Current Repo	67.41	200 epochs, query
Current Repo	67.50	200 epochs, key
Current Repo	70.12	800 epochs, query
Current Repo	70.15	800 epochs, key

3. Reference

• Improved Baselines with Momentum Contrastive Learning [paper] [official code]