This repository is to test machine learning models for reading EMG data to predict different hand positions
To use this repository, you will have to download the data, set up an environment with uv, and setup your wandb login.
To download the data, go to https://www.kaggle.com/datasets/sojanprajapati/emg-signal-for-gesture-recognition/data and download the csv file. Place it in the data/ directory. The EMG-data_description.txt file describes what the data represents, and more information can be found at the Kaggle link
To get uv, please refer to https://docs.astral.sh/uv/getting-started/installation/. After installing, you can just run uv sync in the project folder to get all necessary dependencies. The main code can be run with uv run main.py
In order to log our runs, we are using wandb. For documentation and making an account, go to https://docs.wandb.ai/models/models_quickstart#sign-up-and-create-an-api-key. The first time you run main, you will be promped to create an account or use an existing one. The easiest thing to do is create an account at the link, select use an existing account, and then paste your API key.
There are many things that still need to be done for this project. Here is a list for those that want to contribute but do not know where to get started:
- Improving the CNN: The CNN currently only gets about 50% accuracy on the dataset after 100 epochs, so performance needs to be improved significantly.
- Data cleaning. Since there is a lot of noise and redundancy in the data since it is being sampled every 1ms, I think performance can be drastically improved by doing some kind of low pass filter before inputting to the CNN
- Actual torch Dataloader and Dataset: Right now I just made some functions but ideally we change this to actual torch dataloaders
- Random Forest: So far, we have a CNN, but we also plan on