https://github.com/amruthapatil/nyu-chatbot

Creating simple chatbot using movie scripts from the Cornell Movie Dialogs Corpus based on the PyTorch Chatbot Tutorial.
https://github.com/amruthapatil/nyu-chatbot

high-performance-computing machine-learning pytorch torchscript

Last synced: about 2 months ago
JSON representation

Creating simple chatbot using movie scripts from the Cornell Movie Dialogs Corpus based on the PyTorch Chatbot Tutorial.

• Host: GitHub
• URL: https://github.com/amruthapatil/nyu-chatbot
• Owner: AmruthaPatil
• License: mit
• Created: 2024-04-06T22:40:51.000Z (about 2 years ago)
• Default Branch: main
• Last Pushed: 2024-04-06T23:49:48.000Z (about 2 years ago)
• Last Synced: 2025-06-04T17:56:30.793Z (about 1 year ago)
• Topics: high-performance-computing, machine-learning, pytorch, torchscript
• Language: Jupyter Notebook
• Homepage:
• Size: 1.49 MB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# ChatBot

## Introduction

This project involves training a simple chatbot using movie scripts from the Cornell Movie Dialogs Corpus based on the PyTorch Chatbot Tutorial. The chatbot is implemented using a sequence-to-sequence model with Luong attention mechanism(s). The goal is to create an interactive chatbot that can engage in a meaningful conversation with users.

## Prerequisites

To run this project, you will need the following:

- Python 3.x

- PyTorch

- NumPy

- Matplotlib

- Weights and Biases (W&B)

## Methodology

The project consists of two main parts:

### Part 1: Training a Simple Chatbot

- Data Preparation: The chatbot is trained using the Cornell Movie Dialogs Corpus. The data is preprocessed and prepared for training.

- Model Implementation: Implementation of a sequence-to-sequence model with Luong attention mechanism(s).

- Training and Evaluation: Training and evaluation of the model using mini-batches.

- Decoding Module: Implementation of greedy-search decoding module.

- Interaction: Interaction with the trained chatbot.

- Hyperparameter Tuning: Use of Weights and Biases (W&B) for hyperparameter tuning.

- Profiling: Profiling of the model using PyTorch Profiler to measure the time and memory consumption.

![Training](images/BaiscRun.png)

### Part 2: Transitioning to TorchScript

- TorchScript Conversion: Explanation of differences between tracing and scripting in TorchScript and conversion of the trained model to TorchScript.

- Model Evaluation: Evaluation of the TorchScript model.

- Latency Comparison: Comparison of evaluation latency between the TorchScripted model and the original PyTorch model.

## Results and Observations

- The best hyperparameters were found using W&B hyperparameter tuning, resulting in a minimum loss for the trained model.

![W&B](images/sweepimage.png)

- The hyperparameters that significantly impacted the model's convergence were gradient clipping (clip), teacher forcing ratio (tf_ratio), and learning rate (lr).

![Hyperparameter](images/parameter_loss.jpg)

- Profiling of the model showed the time and memory consumption of the selected model’s operators.

- The transition to TorchScript showed a significant improvement in performance on GPU, with the TorchScript model running approximately 1.93 times faster than the original PyTorch model.

![Table](images/dftable.jpg)

## How to Run and Use the Code

1. Clone this repository.

2. Install the required dependencies.

3. Modify the W&B initialization with appropriate project and user name (Update project="W&BProjectName", entity="W&BUserName")

4. Run the Jupyter notebook `ChatBot.ipynb` to train and evaluate the chatbot model.

5. For hyperparameter tuning, follow the instructions in the notebook to set up and run a sweep using W&B.

6. To convert the trained model to TorchScript and evaluate its performance, follow the instructions in the notebook.

## References

- [PyTorch Chatbot Tutorial](https://pytorch.org/tutorials/beginner/chatbot_tutorial.html)

- [Cornell Movie Dialogs Corpus](https://www.cs.cornell.edu/~cristian/Cornell_Movie-Dialogs_Corpus.html)

- [Weights and Biases Website](https://wandb.ai/site)

- [PyTorch Profiler Example](https://pytorch.org/tutorials/recipes/recipes/profiler_recipe.html)

- [PyTorch JIT and TorchScript](https://towardsdatascience.com/pytorch-jit-and-torchscript-c2a77bac0fff)

Thank you for exploring this repository!