Automatic arm positioning prediction from previous user data via supervised learning.
This project was created to predict the user-selected sensor position as the sensor is swept across a sample. When performing the sweep, only the luminosity reading is recorded. The model's goal is to predict the position a normal user would choose, a continuous value between -1 and 1, in order to potentially automate this process.
The models were trained on around 900 samples, scraped from previous user data, with an 80:20 train test split. The data was augmented by shifting the reading positions to the left and right. Both the luminosity and arm position are normalized between -1 and 1 for better neural network performance. An MLP and ConvNet were applied to the data, but the MLP performed better in both accuracy and efficiency.
Install pytorch and juypterlab with respect to your current system. Then use pip install -r requirements.txt to install the generic requirements. Then install the library with pip install . from the project directory.
manual_train.ipynb and hyperparameter_opt.py can be used to train models. These models can then be used in inference via API's defined in the library component. demo_use.ipynb describes how to use these functions.