| # Curriculum Learning Configuration - MEDIUM MODEL | |
| # Balanced configuration with good performance/memory trade-off | |
| # Model configuration - BALANCED CAPACITY | |
| model: | |
| target_dim: 18 # Will be overridden based on dataset | |
| is_unconditional: 0 | |
| timeemb: 256 # Increased from 128 → 256 (2x larger) | |
| featureemb: 64 # Increased from 16 → 64 (4x larger) | |
| target_strategy: "random" | |
| use_aux_loss: False | |
| aux_weight_order1: 0.4 | |
| aux_weight_order2: 0.4 | |
| aux_loss_normalize: False # CRITICAL FIX: Disable dangerous normalization | |
| aux_loss_max_value: 2.0 | |
| # Training configuration | |
| training: | |
| seed: 1 | |
| batch_size: 256 # Reduced for stability with larger model | |
| validation_split: 0.05 | |
| num_workers: 4 | |
| mask_generation_count: 100 | |
| gradient_clip: 0.5 # Tighter clipping to prevent gradient explosion | |
| weight_decay: 1e-4 # CRITICAL FIX: Proper weight decay for Adam | |
| ratio: 0.7 | |
| checkpoint_save_step: 10 | |
| use_mixed_precision: False # CRITICAL FIX: Disable mixed precision for stability | |
| # Dataset configuration | |
| dataset: | |
| window_length: 100 | |
| split: 10 | |
| mask_ratio: 0.5 | |
| scale_factor: 1 | |
| # Curriculum Learning Parameters | |
| curriculum: | |
| phase1_epochs: 30 | |
| phase2_epochs: 30 | |
| # Phase 1: Easy - Reduced learning rates for gradient stability | |
| mask_ratio_phase1_start: 0.1 | |
| mask_ratio_phase1_end: 0.3 | |
| noise_ratio_phase1_start: 0.0001 | |
| noise_ratio_phase1_end: 0.1 | |
| lr_phase1_start: 2e-4 # CRITICAL FIX: Conservative, stable LR | |
| lr_phase1_end: 2e-4 # CRITICAL FIX: Fixed LR prevents momentum disruption | |
| # Phase 2: Medium - Reduced learning rates for gradient stability | |
| mask_ratio_phase2_start: 0.3 | |
| mask_ratio_phase2_end: 0.6 | |
| noise_ratio_phase2_start: 0.1 | |
| noise_ratio_phase2_end: 0.3 | |
| lr_phase2_start: 2e-4 # Keep consistent with phase 1 | |
| lr_phase2_end: 2e-4 # Keep consistent with phase 1 | |
| # Phase 3: Hard - Reduced learning rates for gradient stability | |
| mask_ratio_phase3_start: 0.6 | |
| mask_ratio_phase3_end: 0.8 | |
| noise_ratio_phase3_start: 0.3 | |
| noise_ratio_phase3_end: 0.5 | |
| lr_phase3_start: 2e-4 # Keep consistent with phase 1 | |
| lr_phase3_end: 2e-4 # Keep consistent with phase 1 | |
| # Diffusion configuration - BALANCED CAPACITY | |
| diffusion: | |
| layers: 6 # Increased from 4 → 6 (1.5x more layers) | |
| channels: 128 # Increased from 64 → 128 (2x larger) | |
| nheads: 8 # Keep at 8 for divisibility (128/8 = 16) | |
| diffusion_embedding_dim: 256 # Increased from 128 → 256 (2x larger) | |
| beta_start: 0.0001 | |
| beta_end: 0.01 | |
| num_steps: 500 | |
| schedule: "linear" | |
| # Expected parameter count: ~2.4M parameters (good balance) | |
| # Fixes applied for gradient stability: | |
| # - Reduced learning rates by 50% across all phases | |
| # - Tighter gradient clipping (1.0 → 0.5) | |
| # - Reduced batch size (384 → 256) for stability | |