Skip to content

BDBC-KG-NLP/ICLR2023_Gradient-Dissipation

Repository files navigation

ICLR2023 Gradient Dissipation

The Offical Code for Our Paper:On The Inadequacy of Optimizing Alignment and Uniformity in Contrastive Learning of Sentence Representations, ICLR2023.

Most of our experiments were done in the changed version of SimCSE's codebase, thanks for their great work!

Environment Configuration

  • Create virtual environment and necessary dependence

     conda create -n temp python=3.7
     pip install -r requirement.txt
  • Install torch 1.12.0 for suitable CUDA version (higher versions should also be compatible)

     # CUDA 11.6
     pip install torch==1.12.0+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
     # CUDA 11.3
     pip install torch==1.12.0+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
     # CUDA 10.2
     pip install torch==1.12.0+cu102 --extra-index-url https://download.pytorch.org/whl/cu102

Data Preparation

Path Modification

  • Change TRAIN_FILE_PATH in run_unsup_example.sh to the path of the pre-trained corpus
  • Change EVAL_FILE_PATH in run_unsup_example.sh to the path of the eval_dataset
  • Change MODEL_PATH in run_unsup_example.sh to the path of the pretrained model
  • Change PATH_TO_SENTEVAL in evaluation.py #17 and trainer.py #85 to the path of SentEval
  • Change PATH_TO_DATA in evaluation.py #18 and trainer.py #86 to the path of the datasets in SentEval

Training and Evaluation

  • replicate the results of the met loss function on BERT-base
bash run_unsup_example.sh
  • Edit the hyper-parameter in the shell to replicate the results of other experiments

Todo List

  • The implementation of all loss functions
  • The codes based on SimCSE
  • The checkpoints in the paper

Citetion

@inproceedings{nie2023inadequacy,
  title={On The Inadequacy of Optimizing Alignment and Uniformity in Contrastive Learning of Sentence Representations},
  author={Nie, Zhijie and Zhang, Richong and Mao, Yongyi},
  booktitle={The Eleventh International Conference on Learning Representations},
  year={2023}
}

About

Code and dataset for our paper "On The Inadequacy of Optimizing Alignment and Uniformity in Contrastive Learning of Sentence Representations", ICLR2023

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published