Hybrid Quantum-Classical Training on Event-based Datasets (TQR)

This repo provides a hybrid training pipeline (Quantum + Classical) for event-based datasets using the Temporal Quantum Representation (TQR).

Supported datasets (processed into .npy TQR tensors):

• N-MNIST (10 classes, 34×34, ATIS)
• N-CARS (2 classes, 120×120, ATIS)
• DVS Gesture (11 classes, 128×128, DVS128)
• e-CK+ (7 classes, 346×260, DAVIS346)

---

🔧 Requirements

Install dependencies:

pip install -r requirements.txt

---

📂 Dataset Structure

Expected directory format:

datasets/processed_data/recognition/
  ├── nmnist_rep_100ms/
  │     ├── 0/  (class folder with .npy TQR tensors)
  │     ├── 1/
  │     └── ...
  ├── ncars_rep_100ms/
  ├── dvsgesture_rep_100ms/
  └── eck+_rep_100ms/

Each .npy file should contain a tensor (H, W, BINS, 2) with binary values:

• [..., 0] → negative polarity
• [..., 1] → positive polarity

---

▶️ Training

Run training for all datasets:

python train.py

Or train a single dataset:

python train.py --dataset nmnist

---

⚙️ Pipeline Overview

1. Load TQR tensors (.npy)
2. Preprocessing: optional spatial pooling, flatten
3. PCA reduction → target dimension = number of qubits
   • N-MNIST → 8
   • N-CARS → 8
   • DVS Gesture → 12
   • e-CK+ → 16
4. Quantum embedding: AngleEmbedding + StronglyEntanglingLayers
5. Classical head: Linear + ReLU + Linear → Softmax

---

📊 Output

• Training/validation accuracy per epoch is printed.
• Best validation accuracy is reported per dataset.

---

🧪 Example

python train.py --dataset ncars --epochs 10 --batch_size 32

=== Dataset: ncars | qubits=8 | classes=2 ===
Epoch 01 | train loss 0.6932 acc 0.51 | val loss 0.6920 acc 0.55 | best 0.55
...