superres_inference.yaml

unets:    
  unet1: # SR
    dim: 64
    num_resnet_blocks: 2
    dim_mults: [1,2,4]
    layer_cross_attns: [False, False, False, True]
    use_linear_attn: False
    layer_attns: False
    attend_at_middle: False
    use_linear_cross_attn: False
    memory_efficient: True
    channels: 1
    attn_heads: 8

superres:
  params:
    image_sizes: [128, 512]
    channels: 1
    timesteps: [25,25]
    condition_on_text: True
    random_crop_sizes: [null, null]
    pred_objectives: v
    lowres_sample_noise_level: 0.1


trainer:
  params:
    split_batches: True
    lr: 5e-4
    dl_tuple_output_keywords_names: ['input_lowres','images', 'texts']

dataset:
  root: ""
  image_size: 512 # image size


dataloader: # This is adapted when launching the training script
  params:
    batch_size: 1
    num_workers: 1
    pin_memory: True
    persistent_workers: True

wandb:
  name_suffix: ""
  project: ""
  entity: ""

checkpoint:
  path: "superres_inference"
  batch_size: 1
  cond_scale: 1.
  save_every_x_it: 200