dnguyencodez/Opt4ML-Semester-Project

iSogCLR PyTorch Implementation

In this repo, we show how to train a self-supervised model by using Global Contrastive Loss (GCL) on a widely used bimodal image-text dataset CC3M.

Getting Started

Try in Colab: https://colab.research.google.com/drive/1FTF-cTcW11Gyrwu8uhTZOXgLsjp49Z9W?usp=sharing

Environment

Setting up a new virtual environment with Conda:

env_name='csce689_proj'
conda create -n "$env_name" python=3.10
conda activate "$env_name"
pip install -r requirements.txt

Training and Evaluation

1. Download the data: cc3m_subset_100k.tar.gz, a 100k subset of the Conceptual Captions dataset; mscoco_val.tar.gz, a 5k subset of the COCO val2014 dataset; clip_train.tar.gz, captions of the previous datasets. The code and data should be structured as follows:

   .
   +--bimodal_exps (code)
   |
   +--clip_train (captions)
   |  +--cc3m_train_subset.json
   |  +--coco_val.json
   |
   +--datasets (images)
   |  +--cc3m_subset_100k
   |  +--mscoco_val
   

2. To train a model on cc3m, use run.slurm if slurm is supported or run

   export PYTHONPATH="$PYTHONPATH:./bimodal_exps"
   export HUGGINGFACE_HUB_CACHE='./checkpoints/huggingface'
   
   data_path=./datasets
   ann_path=./clip_train
   train_image_root=cc3m_subset_100k/
   data=cc3m
   train_file=${data}_train_subset.json
   gamma=0.8
   epochs=30
   
   CUDA_VISIBLE_DEVICES=0 python ./bimodal_exps/clip.py \
       --data_path ${data_path} \
       --ann_path ${ann_path} \
       --train_file ${train_file} \
       --train_image_root ${train_image_root} \
       --output_dir output/isogclr_${data}_g${gamma}_e${epochs} \
       --init_model \
       --use_amp \
       --ita_type sogclr \
       --tau_init 0.01 \
       --sogclr_gamma ${gamma} \
       --eta_init 0.03 --sched cosine \
       --no-distributed \
       --epochs ${epochs}

3. To test the performance of a model on mscoco, use eval.slurm if slurm is supported or run

   export PYTHONPATH="$PYTHONPATH:./bimodal_exps"
   export HUGGINGFACE_HUB_CACHE='./checkpoints/huggingface'
   
   data_path=./datasets
   ann_path=./clip_train
   train_image_root=cc3m_subset_100k/
   data=cc3m
   train_file=${data}_train_subset.json
   gamma=0.8
   epochs=30
   
   CUDA_VISIBLE_DEVICES=0 python ./bimodal_exps/clip.py \
       --data_path ${data_path} \
       --ann_path ${ann_path} \
       --train_file ${train_file} \
       --train_image_root ${train_image_root} \
       --output_dir output/isogclr_${data}_g${gamma}_e${epochs} \
       --init_model \
       --use_amp \
       --ita_type sogclr \
       --tau_init 0.01 \
       --sogclr_gamma ${gamma} \
       --eta_init 0.03 --sched cosine \
       --no-distributed \
       --epochs ${epochs} \
       --evaluate --checkpoint './output/isogclr_cc3m_g0.8_e30/checkpoint_30.pth'

Reference

If you find this tutorial helpful, please cite:

@inproceedings{qiu2023not,
  title={Not All Semantics are Created Equal: Contrastive Self-supervised Learning with Automatic Temperature Individualization},
  author={Zi-Hao Qiu, Quanqi Hu, Zhuoning Yuan, Denny Zhou, Lijun Zhang, and Tianbao Yang},
  booktitle={International Conference on Machine Learning},
  pages={TBD},
  year={2023},
  organization={PMLR}
}