mvclab-ntust-course/course7-llm-lannyyuan-go

course7-llm-lannyyuan-go created by GitHub Classroom

Homework 6 : Learn basic pytorch

• The pytorch tutorial notebook provide several materials to learn Pytorch, Google colab and ipd.
• In the last part of the notebook, there's a simple example to train MLP on CPU.
• There's the screenshot of my week 7 colab training result.

Homework 7 : Learn basic pytorch

README

Overview

This homework involves fine-tuning a DistilBERT model for sequence classification on the IMDB dataset using Hugging Face Transformers library. The training process incorporates LoRA (Low-Rank Adaptation) to enhance the model's performance. The model is trained to classify movie reviews as either positive or negative. Because My Colab GPU quota was used up and I think it's more convenient to run on ipynb on my server, so I don't use Colab to do this part.

Requirements

• Python 3.8+
• PyTorch
• Transformers
• Datasets
• PEFT (Parameter Efficient Fine-Tuning)

Setup

1. Install dependencies:

   pip install torch transformers datasets peft

Dataset

The IMDB dataset is used for training and validation. The dataset is loaded and processed using the datasets library.

Training

Model and Tokenizer

The base model used is distilbert-base-cased. The tokenizer is initialized using the same model checkpoint.

Data Preparation

The dataset is truncated to the first 50 tokens for each review to speed up processing. The dataset is then tokenized and prepared in batches of 16 examples.

DataLoader

DataLoaders are created for training and evaluation datasets.

LoRA Configuration

A LoRA configuration is applied to the model to optimize performance. The configuration includes parameters like rank number, alpha (scaling factor), dropout probability, and target modules.

Training Arguments

The training arguments are defined to control various aspects of the training process such as batch size, learning rate, evaluation strategy, and more.

Trainer

The Hugging Face Trainer is used to manage the training process. It handles the training loop, evaluation, and other functionalities.

Compute Metrics

A custom function compute_metrics is defined to compute the accuracy of the model during evaluation.

Training the Model

The training process is initiated using the trainer.train() method.

Usage

To run the training process, execute the script:

python train.py

Result

• Config 1

  

• Config 2

  

• Config 3

  

Latex version

• The pdf is named as Homework_Latex.pdf, please check.