pip install -r requirements.txtct2echo_pkg/
├── conf/ # Hydra configuration
│ ├── config.yaml # Main config
│ ├── model/dinov3.yaml # Model config
│ ├── training/ # Training hyperparameters
│ └── azureml/ # Azure ML settings
├── ct2echo/ # Core package
│ ├── models/ # DINOv3 + Hypernet
│ ├── training/ # Training logic
│ ├── evaluation/ # Evaluation metrics
│ └── data/ # Data loading
└── scripts/ # Entry points
cd scripts
# Run with default config
python mtl_main_hydra.py
cd scripts
# Dry run (validate config)
python submit_job_hydra.py dry_run=true
# Submit training job
python submit_job_hydra.py
# Debug mode (enables VS Code + JupyterLab)
python submit_job_hydra.py debug_mode=true
Before running on Azure ML, update conf/azureml/production.yaml:
subscription_id: "<YOUR_SUBSCRIPTION_ID>"
resource_group: "<YOUR_RESOURCE_GROUP>"
workspace_name: "<YOUR_WORKSPACE_NAME>"
compute_name: "<YOUR_COMPUTE_NAME>"
environment_name: "<YOUR_ENVIRONMENT_NAME>"
experiment_name: "<YOUR_EXPERIMENT_NAME>"
data_base_path: "Users/<YOUR_USERNAME>/data_weights"| Parameter | Default | Description |
|---|---|---|
model.lora_rank |
16 | LoRA adapter rank |
model.learnable_embedding_dim |
512 | Task embedding dimension |
training.lr |
1e-5 | Learning rate |
training.num_epochs |
20 | Number of epochs |
training.batch_size |
8 | Batch size |
Training outputs are saved to paths.output_dir:
epoch_N/- Checkpoints per epochfinal/- Final model weightshydra_config.yaml- Saved configuration
Pre-computed saliency maps are available on Google Drive.
