| seed: 2022 | |
| data: | |
| target: datasets.cifar10.CIFAR10 | |
| params: | |
| root: ~/data/CIFAR-10/ | |
| img_size: 32 | |
| img_channels: 3 | |
| num_classes: 10 | |
| dataloader: | |
| num_workers: 4 | |
| pin_memory: true | |
| prefetch_factor: 2 | |
| model: | |
| target: models.unet_categorial_adagn.UNetCategorialAdaGN | |
| params: | |
| in_channels: 3 | |
| out_channels: 3 | |
| dim: 128 | |
| dim_mults: | |
| - 1 | |
| - 2 | |
| - 2 | |
| - 2 | |
| use_attn: | |
| - false | |
| - true | |
| - true | |
| - false | |
| num_res_blocks: 2 | |
| num_classes: 10 | |
| attn_head_dims: 64 | |
| resblock_updown: true | |
| dropout: 0.1 | |
| diffusion: | |
| target: diffusions.cfg.ddpm_cfg.DDPMCFG | |
| params: | |
| total_steps: 1000 | |
| beta_schedule: cosine | |
| beta_start: 0.0001 | |
| beta_end: 0.02 | |
| objective: pred_eps | |
| var_type: fixed_large | |
| train: | |
| n_steps: 800000 | |
| batch_size: 128 | |
| micro_batch: 0 | |
| clip_grad_norm: 1.0 | |
| ema_decay: 0.9999 | |
| ema_gradual: true | |
| print_freq: 400 | |
| save_freq: 10000 | |
| sample_freq: 5000 | |
| n_samples_each_class: 10 | |
| p_uncond: 0.2 | |
| optim: | |
| target: torch.optim.AdamW | |
| params: | |
| lr: 0.0002 | |