/pytorch-ssl

pytorch self supervised learning

Primary LanguagePython

Pytorch-SSL

This is pytorch-self-supervised-learning repository.

Table of contents:

  1. tutorial
  2. experimental result

Tutorial

1. how to pre-train your model

  1. clone this repo.

    git clone https://github.com/hankyul2/pytorch-ssl.git
cd pytorch-ssl
pip3 install requirements.txt

  2. run following command.

2. how to fine-tune your model

  1. run following command.

3. how to validate your model

  1. run following command.

    KNN classifier

    # Single-GPU
python3 valid.py -es imagenet1k_knn_224_v1 -ws dino_official
# Multi-GPU 
torchrun --nproc_per_node=4 --master_port=12345 valid.py -es imagenet1k_knn_224_v1 -ws dino_official

    FC classifier

    # Single-GPU
python3 valid.py -es imagenet1k_fc_224_v1 -ws dino_official
# Multi-GPU 
torchrun --nproc_per_node=4 --master_port=12345 valid.py -es imagenet1k_fc_224_v1 -ws dino_official

    Tips

    1. For large sized train dataset, we recommend to run multi-gpu command because, extracting whole features takes too much time in single gpu.
    2. If you have extracted features once, please specify the extracted feature path at feature_path in config/valid.json.
    3. If you don't specify weight settings (-ws), whole weight settings will be used in validation.
    4. If you don't specify model name (-m), whole model in model_weight will be used in validation.