| # Documentary configuration for the indexed stage-2 runner; not csa train. | |
| model: | |
| width: 64 | |
| depth: 8 | |
| attention_chunk_size: 256 | |
| mask_threshold: 0.3125 | |
| data: | |
| train_images: data/train.jsonl | |
| val_images: data/validation.jsonl | |
| train_masks: '' | |
| val_masks: '' | |
| mask_mode: random | |
| random_min_fraction: 0.1 | |
| random_max_fraction: 0.4 | |
| mask_white_is_hole: true | |
| image_size: 256 | |
| center_overlap: 4 | |
| horizontal_flip: true | |
| batch_size: 192 | |
| workers: 32 | |
| train: | |
| output: runs/stage2/run | |
| device: cuda:0 | |
| seed: 2019 | |
| epochs: 120 | |
| decay_start: 20 | |
| lr: 0.0002 | |
| beta1: 0.5 | |
| beta2: 0.999 | |
| reconstruction_weight: 0.1 | |
| lpips_weight: 1.0 | |
| lpips_net: vgg | |
| lpips_batch_size: 8 | |
| consistency_weight: 0.01 | |
| adversarial_weight: 0.002 | |
| discriminator_width: 64 | |
| feature_discriminator_width: 512 | |
| amp: bf16 | |
| log_every: 1 | |
| image_discriminator: patch | |