README.md

FormFit ML Pipeline

This folder contains the training and export pipeline for the current FormFit movement-quality model work.

Contents

• CNN.py Trains a 1D convolutional network on rep-window motion data, evaluates it, prints sample feedback, and exports a Core ML package.

• formfit-data/ Split repetition CSV files used for the current training workflow.

• formfit-labels.csv Labels aligned to the split rep sessions.

• best_model.pth Saved PyTorch checkpoint from training.

• training_history.png Loss curve generated from a training run.

• FormFitModel.mlpackage/ Exported Core ML package for downstream Apple-platform integration.

Install

python3 -m pip install -r requirements.txt

Train + Export

python3 CNN.py

The script is configured to:

• load data relative to its own directory
• save the best PyTorch weights locally
• save the training curve as an image
• export the Core ML package locally

Model Inputs

Each rep window is built from nine motion channels:

• ax, ay, az
• gx, gy, gz
• roll, pitch, yaw

Current Scope

The current model focuses on three movement-quality signals:

• elbow hiking
• shoulder hiking
• torso twist

The pipeline also folds eccentric timing into the reported feedback score during evaluation.