Inference Guideline (Team GhostGrip)

Synapse : The NeuroTech Challenge - Third Place Solution

Check out the results here

This folder contains the trained Synapse-SRW model (SWA) and an inference script to evaluate it on an unseen test dataset with the same structure as the training data.

Folder Contents

GHOSTGRIP_SYNAPSE/
│
├── inference.py                 # Inference script
├── synapse_srw_swa_final.pth    # Trained SWA model
├── norm_stats.json              # Normalization statistics
└── test/                        # (TO BE ADDED BY USER)

---

Test Dataset Structure (IMPORTANT)

After unzipping this submission, place the entire test folder inside this directory.

The test dataset must follow the same structure as training:

test/
├── Session1/
│   ├── session1_subject_1/
│   │   ├── gesture00_trial01.csv
│   │   ├── gesture01_trial02.csv
│   │   └── ...
│   ├── session1_subject_2/
│   │   └── ...
│
├── Session2/
│   └── ...
│
├── Session3/
│   └── ...

• Number of sessions: fixed (3)
• Number of subjects: dynamic
• Number of trials: dynamic
• Number of gestures: 5 (fixed)

No CSV list file is required — the script automatically discovers all subjects and trials.

---

Requirements

• Python 3.9+
• PyTorch
• NumPy
• scikit-learn

(Install dependencies if needed)

pip install torch numpy scikit-learn

---

How to Run Inference

From inside this folder, run:

python inference.py --data_root <path/to/test/dataset>

or (if using a specific Python interpreter):

C:\path\to\python.exe inference.py --data_root <path/to/test/dataset>

---

What the Script Does

1. Loads all CSV trials from the test dataset

2. Applies:

   • Bandpass filtering
   • Normalization using norm_stats.json
   • Sliding window segmentation

3. Runs inference using the SWA-averaged Synapse-SRW model

4. Prints:

   • Accuracy
   • Macro-F1 score
   • Full classification report

Example output:

Trials loaded: 2625
Windowing signals (this may take a few minutes)...
Inference progress: batch 100/800
...
=== FINAL TEST RESULTS ===
Accuracy : 0.9044
Macro-F1 : 0.9044

⚠️ Note: Inference may take a few minutes depending on dataset size and hardware.

---

Notes

• GPU will be used automatically if available.
• No training occurs — this script is inference-only.
• Do not rename model or stats files.