README.md

Text-Style-Transfer

About the Project

Results

Aim

The aim of this project is to utilise various neural network architectures to modify the style of a sentence while preserving its meaning.

Description

Text Style Transfer is the modification of style of a sentence without altering the semantic content. This can be done through a variety of different methods.

Here we have showcased 4 of the methods that we have tried to apply to solve the given problem.

Namely,

1. Transformer-only
2. StyleFormer
3. DeleteRetrieveGenerate
4. CrossAligned

Tech Stack

The techonologies used in this project include:

1. Python
2. Tensorflow
3. Pytorch
4. GoogleColab

File Structure

├── 3b1b-Linear-Algebra-Notes               # Notes made for linear algebra course
    ├── 
    ├── 
├── Coursera-Assignments                    # Coursera Assignments
    ├── Warren-Assignments
    ├── Druhi-Assignments
    ├── Yashvi-Assignments
├── Coursera-Notes                          # Notes on Deep Learning
    ├── Warren-Notes
    ├── Druhi-Notes
    ├── Yashvi-Notes
├── CrossAligned                            # CrossAligned implementation
    ├── code
    ├── data
    ├── img
    ├── tmp
├── DeleteRetrieveGenerate                  # DeleteRetrieve implementation
    ├── data
    ├── src
    ├── tools
    ├── working_dir
├── Mini-Projects                           # Mini projects made
    ├── Deep-Neural-Network-From-Scratch
    ├── IMDB-Lstm
    ├── MNISTDigit
├── Project-Report                          # Report on the project
    ├──  Project-Report.pdf
├── StyleFormer                             # StyleFormer implementation
    ├── data
    ├── evaluator
    ├── models
    ├── outputs
    ├── data.py
    ├── main.py
    ├── train.py
    ├── utils.py
    ├── README.md
├── Transformer                             # Transformer implementation
    ├── data
    ├── weights
    ├── Tokenizer.ipynb
    ├── Transformer.ipynb
    ├── README.md
├── LICENSE                            
├── README.md                          

Contributors

• Druhi Phutane

• Warren Jacinto

• Yashvi Gala

Future Prospects

• The usage of finetuned LLMs such as GPT, PaLM, to achieve the desired fluency in Text Style Transfer.
• Potentially use Reinforcement Learning as described here

Resources and Acknowledgments

• We are gratfeul to CoC VJTI and the Project X program.

• Special thanks to our mentor Labeeb Asari for his experties and guidance throughout the project.

• Additionally, we are also thankful for all the Project X mentors for their inputs and advice on our project.

• Refered the official tensorflow tutorials for understanding the framework.

• Andrew Ng's courses on Coursera for an understanding of neural networks.

• Attention is All You Need for understanding the Transformer architecture.

• Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation for the Style Transformer Implementation.

• Style Transfer from Non-Parallel Text by Cross-Alignment for the CrossAligned implementation

• Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer for the DeleteRetrieve implementation