TensorFlow.js Pose Classification

This project uses TensorFlow.js to create an image classification model to recognize 5 different yoga poses - downdog, goddess, plank, tree, and warrior2.

Overview

The project has the following structure:

• data.js - Loads images from the filesystem and processes them into tensors
• model.js - Defines and trains a CNN model
• main.js - Entry point that loads data and trains the model
• Other helpers and config files

The model is a convolutional neural network (CNN) with the following layers:

• Convolutional
• Max pooling
• Dropout
• Flatten
• Dense

It is trained for 50 epochs with categorical crossentropy loss and adam optimizer.

Usage

Check compatibility:

• I have verified usability on MacOS (M2,2022 chip) and on Ubuntu (Linux-ublts22043).
• Known issues include incompatibilty of latest tfjs node versions with Windows.

To train the model:

1. Clone the repo
2. Install dependencies with npm install
3. Make sure you have training data images in ./DATASET/TRAIN and test data in ./DATASET/TEST
4. Run node main.js (OR) npm run train to execute main.js and train the model

The trained model and its weights are saved in the repo itself.

Development

Please commit to dev-primary branch to suggest any changes.

Some ways the project can be improved:

• Add data augmentation
• Use transfer learning from a pretrained model
• Deploy the model to make predictions on new images

Author

Email: arghya[at]nyu[dot]edu