Text-Assisted Image Manipulation - GAN
In this project we explore the idea of changing the images according to the textual captions through generative networks
There's two primary ways in which you can use our project: use our publicly deployed version of TAIM-GAN or deploy it locally on your machine. Also, feel free to fork and modify the source code for latter reseach
Visit the live demo of our project on Hugging Face! It has some interesting examples:
To set up the project locally through docker, just do the following:
- Clone our repository using
git clone https://github.com/Dmmc123/taim-gan.git - Deploy the Gradio web application using
docker-compose up
We used three datasets for training, evaluation, and deployment of TAIM-GAN:
Example of training script:
train.py --data-dir data --split train --num-capt 5 --batch-size 16 --num-workers 8
Example of evaluation script:
compute_metrics.py --data-dir data --split test --num-capt 10 --batch-size 32 --num-workers 4
For additonal information about possible values of arguments and their meaning, please type train.py --help or compute_metrics.py --help
Project uses template from cookiecutter for data science projects
├── notebooks <- Jupyter Notebooks for interim results of coding tasks
│
├── references <- Main papers we used as a source
│
├── src <- Source files of the project
| |
│ ├── data <- Code for preparing the datasets and vectorizing texts
| | |
│ | ├── stubs <- Folder for examples in the Gradio application
| | |
│ | ├── collate.py <- Function for preprocessing output of datasets for inference
│ | ├── datasets.py <- Code for processing the raw data from datasets
│ | └── tokenizer.py <- API for tokenizing text captions
| |
│ ├── models <- Code for model components and inference utility functions
| | |
| | ├── modules <- All the atomic modules that constitute TAIM-GAN
| | |
│ | ├── losses.py <- Loss functions for Discriminator and Generator
│ | └── utils.py <- Helper functions for loading/dumping weights of models and saving plots
| |
| └── config.py <- Constants used throughout the project
|
├── tests <- Code for integration and unit testing
| │
| ├── integration <- Tests for checking the whole flow of TAIM-GAN
| │
| └── unit <- Scripts for testing the mudules in atomic way
|
├── app.py <- Code for Gradio web application
├── compute_metrics.py <- Code for computing Inceptions Scores on datasets
├── docker-compose.yaml <- Docker config for deploying the Gradio web app locally
├── requirements.txt <- List of all dependencies of the project
└── train.py <- Code for training the model
The project is primarily based on LWGAN research. Here is what we did differently:
- Removed the redundant modules from the source code
- Refactored from scratch the existing code by providing type hints for the most part of code-base
- Covered most parts of code with unit and intergrations tests
- Removed the deprecated functionality and upgraded the project according the latest version of PyTorch 1.12.1
- Collected new dataset with captions for facial pictures and finetuned TAIM-GAN with it
Also here are some other researches we used as an additional reference: