Saama Word Guesser ❓

This project mimics the problem with hangman/dumb charades like NLP word games. This NLP model will predict the correct word from an intentionally obscured word based on its description.

Example:

Input masked word = DEM_G_A_HY and

Description = is the statistical study of populations, especially human beings.

Model prediction/output = DEMOGRAPHY

Method to Execute the Project

1. Install the libraries using the command - pip install -r requirements.txt
2. Run the app_gradio.py file - python3 app_gradio.py

Training a model from Scratch

1. Run python3 train.py --dataset_csv_path path/to/dataset.csv, the csv must contain words in the first column, and their meanings in the second column.

For example, run python3 train.py --dataset_csv_path ./data/train.csv to train on English words (dataset included).

OR

1. Build docker image with python3 build_container.py train
2. Run sudo docker run train path/to/dataset.csv

Gradio app

1. Run python3 app_gradio.py to launch gradio-based web-app.

OR

1. Build docker image with python3 build_container.py inference_gradio
2. Run sudo docker run inference_gradio

Rest API

1. Run python3 app_rest.py to launch flask REST server.

OR

1. Build docker image with python3 build_container.py inference_rest
2. Run sudo docker run inference_rest

Example request

curl --header "Content-Type: application/json" \
--request POST \
--data '{"masked_word":"DEM_G_A_HY","description":"the statistical study of populations."}' \
http://127.0.0.1:5000/guess_word

Project Details:

• train.py is the training script, uses PyTorch-Lightning, pass --dataset_csv_path argument to the script (works for the docker container as well).
• tests contains unit tests for modules in core
• core contains implementations of the dataset, model, tokenizer etc.
• data contains csvs for training, testing, dumped tokenizer etc.
• build_tokenizer.py builds a character-level tokenizer for a given CSV.
• clean_data.py is used for cleaning the open-source dataset used.
• app_gradio.py is a simple gradio inference for inference with a trained model (included).
• app_rest.py is a simple REST API (flask).
• dockerization_app contain Dockerfiles for inference and training.
• build_container.py builds docker containers for inference_gradio, inference_rest, and inference_training. Example: python3 build_container.py inference_gradio.
• logs contain the chekpoint of model.