PyTorch trainer and model for Sequence Classification

Last update: Dec 09, 2022

Overview

PyTorch-trainer-and-model-for-Sequence-Classification

After cloning the repository, modify your training data so that the training data is a .csv file and it has 2 columns: Text and Label

In the below example, we will assume that our training data has 3 labels, the name of our training data file is train_data.csv

Example Usage

Import dependencies

import pandas as pd
import numpy as np
from transformers import AutoModel, AutoTokenizer, AutoConfig

from EarlyStopping import *
from modelling import *
from utils import *

Specify arguments

args.pretrained_path will be the path of our pretrained language model

class args:
    fold = 0
    pretrained_path = 'bert-base-uncased'
    max_length = 400
    train_batch_size = 16
    val_batch_size = 64
    epochs = 5
    learning_rate = 1e-5
    accumulation_steps = 2
    num_splits = 5

Create train and validation data

In this example we will train the model using cross-validation. We will split our training data into args.num_splits folds.

df = pd.read_csv('./train_data.csv')
df = create_k_folds(df, args.num_splits)

df_train = df[df['kfold'] == args.fold].reset_index(drop = True)
df_valid = df[df['kfold'] == args.fold].reset_index(drop = True)

Load the language model and its tokenizer

config = AutoConfig.from_pretrained(args.path)
tokenizer = AutoTokenizer.from_pretrained(args.path)
model_transformer = AutoModel.from_pretrained(args.path)

Prepare train and validation dataloaders

features = []
for i in range(len(df_train)):
    features.append(prepare_features(tokenizer, df_train.iloc[i, :].to_dict(), args.max_length))
    
train_dataset = CreateDataset(features)
train_dataloader = create_dataloader(train_dataset, args.train_batch_size, 'train')

features = []
for i in range(len(df_valid)):
    features.append(prepare_features(tokenizer, df_valid.iloc[i, :].to_dict(), args.max_length))
    
val_dataset = CreateDataset(features)
val_dataloader = create_dataloader(val_dataset, args.val_batch_size, 'val')

Use EarlyStopping and customize the score function

NOTE: The customized score function should have 2 parameters: the logits, and the actual label

def accuracy(logits, labels):
    logits = logits.detach().cpu().numpy()
    labels = labels.detach().cpu().numpy()
    pred_classes = np.argmax(logits * (1 / np.sum(logits, axis = -1)).reshape(logits.shape[0], 1), axis = -1)
    pred_classes = pred_classes.reshape(labels.shape)
    
    return np.sum(pred_classes == labels) / labels.shape[0]

es = EarlyStopping(mode = 'max', patience = 3, monitor = 'val_acc', out_path = 'model.bin')
es.monitor_score_function = accuracy

Create and train the model

Calling the fit method, the training process will begin

model = Model(config, model_transformer, num_labels = 3)
model.to('cuda')
num_train_steps = int(len(train_dataset) / args.train_batch_size * args.epochs)
model.fit(args.epochs, args.learning_rate, num_train_steps, args.accumulation_steps, 
          train_dataloader, val_dataloader, es)

NOTE: To complete the cross-validation training process, run the code above again with args.fold equals 1, 2, ..., args.num_splits - 1

PyTorch trainer and model for Sequence Classification

Related tags

Overview

PyTorch-trainer-and-model-for-Sequence-Classification

Example Usage

Import dependencies

Specify arguments

Create train and validation data

Load the language model and its tokenizer

Prepare train and validation dataloaders

Use EarlyStopping and customize the score function

Create and train the model

Owner

NhanTieu

Conversational text Analysis using various NLP techniques

Train Yolov4 using NBX-Jobs

Open-L2O: A Comprehensive and Reproducible Benchmark for Learning to Optimize Algorithms

AquaTimer - Programmable Timer for Aquariums based on ATtiny414/814/1614

Code for the paper: Learning Adversarially Robust Representations via Worst-Case Mutual Information Maximization (https://arxiv.org/abs/2002.11798)

An AI Assistant More Than a Toolkit

Official Code for "Non-deep Networks"

Reverse engineer your pytorch vision models, in style

A hifiasm fork for metagenome assembly using Hifi reads.

WormMovementSimulation - 3D Simulation of Worm Body Movement with Neurons attached to its body

The Balloon Learning Environment - flying stratospheric balloons with deep reinforcement learning.

SegTransVAE: Hybrid CNN - Transformer with Regularization for medical image segmentation

Unofficial TensorFlow implementation of the Keyword Spotting Transformer model

Official Pytorch implementation for "End2End Occluded Face Recognition by Masking Corrupted Features, TPAMI 2021"

Deep Learning Theory

The code of paper 'Learning to Aggregate and Personalize 3D Face from In-the-Wild Photo Collection'

[ACL 20] Probing Linguistic Features of Sentence-level Representations in Neural Relation Extraction

Bulk2Space is a spatial deconvolution method based on deep learning frameworks

Offcial implementation of "A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Prediction, ICCV-2021".

torchbearer: A model fitting library for PyTorch