KoSimCSE Training on Amazon SageMaker

Introduction

SimCSE is a highly efficient and innovative embedding technique based on the concept of contrastive learning. Unsupervised learning can be performed without the need to prepare ground-truth labels, and high-performance supervised learning can be performed if a good NLI (Natural Language Inference) dataset is prepared. The concept is very simple and the psudeo-code is intuitive, so the implementation is not difficult, but I have seen many people still struggle to train this model.

The official implementation code from the authors of the paper is publicly available, but it is not suitable for a step-by-step implementation. Therefore, we have reorganized the code based on Simple-SIMCSE's GitHub so that even ML beginners can train the model from the scratch with a step-by-step implementation. It's minimalist code for beginners, but data scientists and ML engineers can also make good use of it.

Added over Simple-SimCSE

• Added the Supervised Learning part, which shows you step-by-step how to construct the training dataset.
• Added Distributed Learning Logic. If you have a multi-GPU setup, you can train faster.
• Added SageMaker Training. ml.g4dn.xlarge trains well, but we recommend ml.g4dn.12xlarge or ml.g5.12xlarge for faster training.

Requirements

We recommend preparing an Amazon SageMaker instance with the specifications below to perform this hands-on.

SageMaker Notebook instance

• ml.g4dn.xlarge

SageMaker Training instance

• ml.g4dn.xlarge (Minimum)
• ml.g5.12xlarge (Recommended)

Datasets

For supervised learning, you need an NLI dataset that specifies the relationship between the two sentences. For unsupervised learning, we recommend using wikipedia raw data separated into sentences. This hands-on uses the dataset registered with huggingface, but you can also configure your own dataset.

The datasets used in this hands-on are as follows

Supervised

• Klue-NLI
• Kor-NLI

Unsupervised

• kowiki-sentences: Data from 20221001 Korean wiki split into sentences using kss (backend=mecab) morphological analyzer.

How to Train

JupyterLab / Jupyter Notebook / Colab

Supervised

• A.1_sup-prepare-nli-dataset.ipynb: Prepare Dataset for training
• A.2_sup-train-dev.ipynb: Training on Local Environment
• A.3_sm-sup-train.ipynb: Training on SageMaker

Unsupervised

• B.2_unsup-train-dev.ipynb: Training on Local Environment
• B.3_sm-unsup-train.ipynb: Training on SageMaker

Command line (Local)

Supervised

cd src
python sup_prepare_dataset.py
bash sup_run_local.sh

Unsupervised

cd src
bash unsup_run_local.sh

Inference example

from transformers import AutoModel, AutoTokenizer
from serving_src.model import SimCSEModel
model = SimCSEModel.from_pretrained("daekeun-ml/KoSimCSE-supervised-roberta-large") 
tokenizer = AutoTokenizer.from_pretrained("daekeun-ml/KoSimCSE-supervised-roberta-large") 

from src.infer import show_embedding_score
sentences = ['이번 주 일요일에 분당 이마트 점은 문을 여나요?',
             '일요일에 분당 이마트는 문 열어요?',
             '분당 이마트 점은 토요일에 몇 시까지 하나요']
show_embedding_score(tokenizer, model.cpu(), sentences)

Performance

We trained with parameters similar to those in the paper and did not perform any parameter tuning. Higher max sequence length does not guarantee higher performance; building a good NLI dataset is more important

{
  "batch_size": 64,
  "num_epochs": 1 (for unsupervised training), 3 (for supervised training)
  "lr": 1e-05 (for unsupervised training), 3e-05 (for supervised training)
  "num_warmup_steps": 0,
  "temperature": 0.05,
  "lr_scheduler_type": "linear",
  "max_seq_len": 32,
  "use_fp16": "True",
}

KLUE-STS

Kor-STS

References

• Simple-SimCSE: https://github.com/hppRC/simple-simcse
• KoSimCSE: https://github.com/BM-K/KoSimCSE-SKT
• SimCSE (official): https://github.com/princeton-nlp/SimCSE
• SimCSE paper: https://aclanthology.org/2021.emnlp-main.552

License Summary

This sample code is provided under the MIT-0 license. See the license file.