project:
  root: .
data:
  coco_root: None
outputs:
  root: outputs
  vae_dir: outputs/vae
  ldm_dir: outputs/ldm
  latent_dir: outputs/latents
  sample_dir: outputs/samples
  log_dir: logs
cache:
  root: cache
dataset:
  name: coco_captions
  root_key: data.coco_root
  train_split: train2017
  val_split: val2017
  resolution: 256
  num_workers: 8
  pin_memory: true
model:
  name: autoencoder_kl
  in_channels: 3
  out_channels: 3
  latent_channels: 8
  base_channels: 128
  channel_multipliers:
  - 1
  - 2
  - 4
  - 4
  num_res_blocks: 3
  dropout: 0.0
  use_attention: true
  attention_heads: 4
  scaling_factor: 1.032
attention_resolutions:
- 32
experiment:
  name: vae_coco_256_small
train:
  seed: 42
  precision: bf16
  batch_size: 16
  gradient_accumulation_steps: 4
  num_workers: 8
  max_epochs: 100
  lr: 0.0001
  weight_decay: 0.0
  betas:
  - 0.9
  - 0.999
  grad_clip: 1.0
  log_every: 100
  validate_every: 1
  save_every: 1
  sample_every: 1
  num_sample_images: 8
  output_dir_key: outputs.vae_dir
  initialize_from_scratch: true
  resume_from: outputs/vae/vae_coco_256_small/checkpoints/last.pt
  finetune_from: null
early_stopping:
  enabled: true
  patience: 15
  min_delta: 0.0
  monitor_metric: val_total_loss
loss:
  recon_loss_type: l1
  recon_weight: 1.0
  use_lpips: true
  lpips_net: vgg
  perceptual_weight: 0.1
  kl_weight: 1.0e-06
  kl_warmup_steps: 10000
optimizer:
  name: adamw