| # Base diffusion configuration | |
| # Provides default values for diffusion training and inference | |
| # Device and compilation settings | |
| compile: false | |
| device: cpu | |
| precision: 'fp32' | |
| seed: 42 | |
| checkpoint_path: '' | |
| # Model configuration defaults | |
| model: | |
| model: 'DiT-S' # Default to smaller model | |
| noise_schedule: 'linear' # Default noise schedule | |
| l1_loss: false # Use MSE loss by default | |
| diffusion_steps: 1000 # Standard diffusion steps | |
| max_diffusion_step: 1000 # Full diffusion by default | |
| context_size: 128 # Default context size | |
| # Data configuration defaults | |
| data: | |
| train_dataset_path: '' # Must be specified in child configs | |
| start: 0 | |
| end: 1000 # Small default dataset | |
| shuffle: true | |
| seq_len: 64 # Default sequence length | |
| stride: 8 # Default stride | |
| cycle_length: 32 # Default cycle length | |
| beatmap_class: false # Disable by default | |
| difficulty_class: false # Disable by default | |
| mapper_class: false # Disable by default | |
| descriptor_class: false # Disable by default | |
| circle_size_class: false # Disable by default | |
| class_dropout_prob: 0.1 # Conservative dropout | |
| diff_dropout_prob: 0.1 | |
| mapper_dropout_prob: 0.1 | |
| descriptor_dropout_prob: 0.1 | |
| cs_dropout_prob: 0.1 | |
| descriptors_path: '' # Must be specified in child configs | |
| mappers_path: '' # Must be specified in child configs | |
| num_diff_classes: 10 # Default number of classes | |
| max_diff: 10 # Default max difficulty | |
| num_cs_classes: 10 # Default number of CS classes | |
| double_time_prob: 0.0 # Disable by default | |
| distance_std: 0.05 # Conservative noise | |
| # Dataloader defaults | |
| dataloader: | |
| num_workers: 2 # Conservative default | |
| # Optimizer defaults | |
| optim: | |
| name: adamw | |
| base_lr: 1e-4 # Conservative learning rate | |
| batch_size: 32 # Small default batch size | |
| total_steps: 100000 # Default training steps | |
| warmup_steps: 5000 # Default warmup | |
| lr_scheduler: linear # Simple scheduler | |
| weight_decay: 0.01 # Light regularization | |
| grad_acc: 1 # No gradient accumulation by default | |
| gradient_clip: 1.0 # Gradient clipping | |
| final_cosine: 1e-6 # Final learning rate | |
| # Checkpoint defaults | |
| checkpoint: | |
| every_steps: 10000 # Save every 10k steps | |
| # Logging defaults | |
| logging: | |
| log_with: 'tensorboard' # Default to tensorboard | |
| every_steps: 100 # Log every 100 steps | |
| mode: 'offline' # Offline by default | |
| # Hydra defaults | |
| hydra: | |
| job: | |
| chdir: false # Don't change directory by default | |