make_a_scene/img_config.yaml

mode: pretrain_image
devices: 
  - 0
  - 1
  - 2
  - 3
  - 4
  - 5
  - 6
    #- 7

total_steps: 4700000
accumulate_grad: 5
resume: True
#checkpoint: /home/ubuntu/Make-A-Scene/outputs/pretrain_image/2022-06-10/06-36-07/checkpoint.pt
checkpoint: /home/ubuntu/Make-A-Scene/outputs/pretrain_image/2022-06-13/22-05-22/checkpoint_21.0.pt
log_period: 50
save_period: 1000
batch_size: 3  # 192 for 256 model and 128 for 512 model

model:
  _target_: models.VQBASE
  embed_dim: 256
  n_embed: 8192
  init_steps: 3000
  reservoir_size: 12500 # 2e5 / 8
  ddconfig:
    z_channels: 256
    in_channels: 3
    out_channels: 3
    channels: [128, 128, 128, 256, 512, 512]  # [1, 1, 2, 4, 4]
    num_res_blocks: 2
    resolution: 512
    attn_resolutions:
    - 32
    dropout: 0.0

optimizer:
  vq:
    lr: 5e-6
    betas:
      - 0.5
      - 0.9
  disc:
    lr: 4.5e-6
    betas:
      - 0.5
      - 0.9

dataset:
  _target_: Data.dataset_preprocessor_web.S3ProcessedDataset
  resampled: True
  names:
    - cc3m
    - cc12m
#  path: file:D:/PycharmProjects/Make-A-Scene/server/Make-A-Scene/dataset/coco/{00000..00004}.tar
#  path: file:D:/PycharmProjects/Make-A-Scene/server/Make-A-Scene/dataset/coco/great_dataset.tar

loss:
  #_target_: losses.VQVAEWithBCELoss
  _target_: losses.loss_img.VQLPIPSWithDiscriminator
  disc_start: 250001
  disc_weight: 0.8
  codebook_weight: 1.0

dataloader:
  batch_size: ${batch_size}
  num_workers: 8
  pin_memory: True

hydra:
  job:
    chdir: True
  run:
    dir: ./outputs/${mode}/${now:%Y-%m-%d}/${now:%H-%M-%S}