System PSU v2

SensaLimb

Smart arm prosthesis project.

ML Model (Latest Recommended): Toro Ossaba Conv1D

Use this model family as the current default:

• Model family folder: ML/models/conv1d_angle_toro_ossaba
• Dataset folder (required): ML/datasets/Toro Ossaba
• Training script: ML/train/train_conv1d_angle_toro_ossaba.py

1) Install ML dependencies

From the repository root (PowerShell):

python -m pip install --upgrade pip
python -m pip install -r requirements-ml.txt

2) Download and extract the Toro Ossaba dataset (required)

Dataset source:

• https://zenodo.org/records/7946782

Extract location must be exactly:

• ML/datasets/Toro Ossaba

Important extraction rule:

• After extraction, your .txt dataset files must be inside ML/datasets/Toro Ossaba/... (subject subfolders).
• Do not leave the files nested one level too deep (for example, avoid ML/datasets/Toro Ossaba/Toro Ossaba/...).

Quick check (PowerShell):

Get-ChildItem "ML/datasets/Toro Ossaba" -Recurse -Filter *.txt | Select-Object -First 5 FullName

3) Install/extract the Toro Ossaba trained model files

If you received model files as a zip/archive, extract them into:

• ML/models/conv1d_angle_toro_ossaba

Expected model artifacts in the selected run folder:

• model.keras
• model.tflite
• model_int8.tflite (optional but recommended for embedded use)
• meta.json

Example expected path pattern:

• ML/models/conv1d_angle_toro_ossaba/<run_name>/model.tflite

If your extraction creates an extra folder level, move files so that the run folder directly contains meta.json and model files.

4) Run the model (inference/validation)

Pick your run folder (example):

$MODEL_DIR = "ML/models/conv1d_angle_toro_ossaba/seed9_20260117_034451_884"

A. Compare float vs int8 model outputs

python ML/tools/compare_tflite_models.py --model-dir "$MODEL_DIR" --input-root "ML/datasets/Toro Ossaba" --samples 50

B. Generate prediction-vs-true plots

python ML/tools/plot_tflite_pred_vs_true.py --model-dir "$MODEL_DIR" --input-root "ML/datasets/Toro Ossaba" --samples 2000

Output images are saved in the same run folder:

• pred_vs_true_float.png
• pred_vs_true_int8.png

5) (Optional) Train a new Toro Ossaba model run

python ML/train/train_conv1d_angle_toro_ossaba.py --use-file-split --movement all --emg-source filtered --angle-history-sec 0.1 --tflite-int8 --seed 9 --run-name seed9_new

This creates a new run folder under ML/models/conv1d_angle_toro_ossaba.

Common setup mistakes

• Dataset extracted to wrong folder (must be ML/datasets/Toro Ossaba).
• Extra nested folder after extraction.
• Model archive extracted to wrong location.
• Missing model.tflite / meta.json in selected run folder.