bibychacko / asl_alphabet_classifier

Final project for MAIS 202 bootcamp // classifier for ASL alphabet hand gestures

ASL_alphabet_classifier

Final project for McGill AI Society Intro to ML Bootcamp (Winter 2019).

Training data retrieved from Kaggle.

Project description

This ASL Alphabet Classifier project is a web app that classifies American Sign Language gestures contained in an input image. I built the classification model using Pytorch and the web app's backend using Flask. Furthermore, I retrieved training data from Kaggle, performed data augmentation with OpenCV and torchvision, and utilized a vgg-11 CNN architecture.

Running the app

To run the web app, install all packages in requirements.txt. Then, change into the main directory of this repository and run

python web_app.py

Lastly, open a browser and navigate to your http://localhost:5000.

Repository organization

This repository contains the scripts used to both train the model and build the web app.

1. deliverables/
   • deliverables submitted to the MAIS Intro to ML Bootcamp organizers
2. model/
   • Data, final model (results), and Python scripts used to train the VGG model (I also attempted an SVM; the scripts for that are in this folder too)
   • CNN_vgg_gpu.py contains the Pytorch Dataloader class used for training
   • get_confusion_matrix.py loads the trained model and tests its performance on the validation set and personal photos
   • train_CNN_gpu.py is a script to train the model on Microsoft Azure's DSVM (with cuda enabled)
3. static/
   • CSS and javascript files for landing page
4. templates/
   • HTML template for landing page
5. predictor.py
   • python class to store model and its methods for web app
6. web_app.py
   • main python script to instantiate Flask server