KoSimCSE Training on Amazon SageMaker

Usage

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from transformers import AutoConfig, PretrainedConfig, PreTrainedModel
from transformers import AutoModel, AutoTokenizer, logging

class SimCSEConfig(PretrainedConfig):
    def __init__(self, version=1.0, **kwargs):
        self.version = version
        super().__init__(**kwargs)

class SimCSEModel(PreTrainedModel):
    config_class = SimCSEConfig

    def __init__(self, config):
        super().__init__(config)
        self.backbone = AutoModel.from_pretrained(config.base_model)
        self.hidden_size: int = self.backbone.config.hidden_size
        self.dense = nn.Linear(self.hidden_size, self.hidden_size)
        self.activation = nn.Tanh()

    def forward(
        self,
        input_ids: Tensor,
        attention_mask: Tensor = None,
        # RoBERTa variants don't have token_type_ids, so this argument is optional
        token_type_ids: Tensor = None,
    ) -> Tensor:
        # shape of input_ids: (batch_size, seq_len)
        # shape of attention_mask: (batch_size, seq_len)
        outputs: BaseModelOutputWithPoolingAndCrossAttentions = self.backbone(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
        )

        emb = outputs.last_hidden_state[:, 0]

        if self.training:
            emb = self.dense(emb)
            emb = self.activation(emb)

        return emb
    
def show_embedding_score(tokenizer, model, sentences):
    inputs = tokenizer(sentences, padding=True, truncation=True, return_tensors="pt")
    embeddings = model(**inputs)
    score01 = cal_score(embeddings[0,:], embeddings[1,:])
    score02 = cal_score(embeddings[0,:], embeddings[2,:])
    print(score01, score02)

def cal_score(a, b):
    if len(a.shape) == 1: a = a.unsqueeze(0)
    if len(b.shape) == 1: b = b.unsqueeze(0)
    a_norm = a / a.norm(dim=1)[:, None]
    b_norm = b / b.norm(dim=1)[:, None]
    return torch.mm(a_norm, b_norm.transpose(0, 1)) * 100 

# Load pre-trained model
model = SimCSEModel.from_pretrained("daekeun-ml/KoSimCSE-supervised-roberta-large")
tokenizer = AutoTokenizer.from_pretrained("daekeun-ml/KoSimCSE-supervised-roberta-large")

# Inference example
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)

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

Requirements

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

SageMaker Notebook instance

SageMaker Training instance

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

Unsupervised

How to train

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": 3e-05,
  "num_warmup_steps": 0,
  "temperature": 0.05,
  "lr_scheduler_type": "linear",
  "max_seq_len": 32,
  "use_fp16": "True",
}

KLUE-STS

Model Avg Cosine Pearson Cosine Spearman Euclidean Pearson Euclidean Spearman Manhattan Pearson Manhattan Spearman Dot Pearson Dot Spearman
KoSimCSE-RoBERTa-base (Unsupervised) 81.17 81.27 80.96 81.70 80.97 81.63 80.89 81.12 80.81
KoSimCSE-RoBERTa-base (Supervised) 84.19 83.04 84.46 84.97 84.50 84.95 84.45 82.88 84.28
KoSimCSE-RoBERTa-large (Unsupervised) 81.96 82.09 81.71 82.45 81.73 82.42 81.69 81.98 81.58
KoSimCSE-RoBERTa-large (Supervised) 85.37 84.38 85.99 85.97 85.81 86.00 85.79 83.87 85.15

Kor-STS

Model Avg Cosine Pearson Cosine Spearman Euclidean Pearson Euclidean Spearman Manhattan Pearson Manhattan Spearman Dot Pearson Dot Spearman
KoSimCSE-RoBERTa-base (Unsupervised) 81.20 81.53 81.17 80.89 81.20 80.93 81.22 81.48 81.14
KoSimCSE-RoBERTa-base (Supervised) 85.33 85.16 85.46 85.37 85.45 85.31 85.37 85.13 85.41
KoSimCSE-RoBERTa-large (Unsupervised) 81.71 82.10 81.78 81.12 81.78 81.15 81.80 82.15 81.80
KoSimCSE-RoBERTa-large (Supervised) 85.54 85.41 85.78 85.18 85.51 85.26 85.61 85.70 85.90

References