QCM

Quantum advantage for medical applications.

Quick Start

1. Create a virtual environment:

   python -m venv venv
   source venv/bin/activate

2. Install dependencies:

   pip install -r requirements.txt

3. Set up the data: Download and extract the dataset into the src/data directory.

   wget "https://surfdrive.surf.nl/index.php/s/mx0RdFXHHDfcY3g/download" -O pcam_tcga.tar
   tar -xf pcam_tcga.tar -C src/data/

4. Train the model: Choose between classical and quantum mode by changing the --mode flag.

   • Classical Mode:

     python src/train.py --config ./configs/config.yaml --mode classical

   • Quantum Mode:

     python src/train.py --config ./configs/config.yaml --mode quantum

Configuration (configs/config.yaml)

The training behavior is controlled by config.yaml. Here is a brief overview of the key parameters:

• dataset_type: Toggles between 'pcam' (binary image classification) and 'tcga' (multiclass embedding classification).
• model:
  • latent_dim: The size of the embedding vector produced by the backbone.
  • n_qubits: The number of qubits to use in the quantum head.
  • n_quantum_layers: The number of repeated layers in the quantum circuit's ansatz.
  • num_classes: Number of unique classes for the TCGA dataset.
  • quantum_head_type: The type of quantum embedding to use, either 'amplitude' or 'angle'.
  • entangling_layer: The type of entangling layer in the quantum circuit, either 'strong' or 'basic'.
• training:
  • batch_size: The number of samples per batch.
  • lr: The learning rate for the Adam optimizer.
  • epochs: The total number of training epochs.

Expected Outputs

After running the training script, you can expect the following outputs:

• Trained Model: The trained model backbone will be saved in the models/ directory (e.g., models/model_backbone_tcga_classical_20231027_103000.pt).
• t-SNE Visualization: A t-SNE visualization of the latent space will be saved in the models/ directory (e.g., models/latent_tsne_tcga_classical_20231027_103000.png).
• Training Logs: Training and validation losses for each run are logged in logs/training_log.csv.
• TensorBoard Logs: You can view live training metrics by running:

  tensorboard --logdir runs