| # Conditional diffusion model configuration (shared across encoders/areas). | |
| # Single-channel 64x64 DDPM conditioned on a 50-d response vector. | |
| unet: | |
| image_size: 64 | |
| in_channels: 1 | |
| out_channels: 1 | |
| block_channels: [64, 128, 256, 512] | |
| layers_per_block: 2 | |
| cross_attention_dim: 256 | |
| conditioning: | |
| embedding_dim: 256 | |
| mlp_layers: [50, 256, 256, 256] # 50 neurons → 256-d token | |
| mlp_activation: silu | |
| cfg_dropout_rate: 0.1 # classifier-free guidance dropout | |
| observation_noise_sigma: 0.22 # σ for ξ added to r at training time | |
| scheduler: | |
| num_train_timesteps: 1000 | |
| beta_schedule: linear | |
| beta_start: 1.0e-4 | |
| beta_end: 0.02 | |
| prediction_type: epsilon | |
| train: | |
| data: tiny_imagenet | |
| data_split: train | |
| n_train_images: 100000 | |
| optimizer: adamw | |
| lr: 1.0e-4 | |
| weight_decay: 1.0e-4 | |
| scheduler: cosine | |
| warmup_steps: 500 | |
| batch_size: 128 | |
| epochs: 200 # 200-ep checkpoint used for figures | |
| precision: bfloat16 | |
| gradient_clip: 1.0 | |
| seed: 42 | |