run_dir: runs/exp004
seed: 4
device: cuda

performance:
  flash_sdp: true
  mem_efficient_sdp: true
  math_sdp: true
  torch_compile: true

data:
  roots:
    - "E:/FrameNN_subset"
  manifest: "runs/exp004/dataset_manifest.jsonl"
  manifest_errors_log: "runs/exp004/manifest_errors.jsonl"
  manifest_force_count: false
  backend: "decord"
  format: "wds"
  npz_dir: ""
  npz_manifest: ""
  wds_urls: "file:/workspace/wds/shard-{000000..000150}.tar"
  wds_shuffle: 1000
  wds_image_ext: "png"
  max_total_tries: 500
  extensions: [".mp4", ".mkv", ".avi", ".webm", ".flv", ".mov", ".mpg", ".mpeg"]
  train_crop: [384, 384]
  min_size: [384, 384]
  resize_short_edge: 0
  frame_stride_min: 1
  frame_stride_max: 3
  reader_cache: 1
  max_tries: 10
  random_horizontal_flip: true

loader:
  batch_size: 16
  num_workers: 10
  pin_memory: true
  persistent_workers: true
  prefetch_factor: 4

train:
  epochs: 11
  steps_per_epoch: 9437
  amp: "bf16" #bf16
  grad_accum_steps: 2
  clip_grad_norm: 0.5
  log_interval: 50
  save_interval_epochs: 1
  resume: ""
  profile: false
  profile_interval: 20
  profile_warmup_steps: 10
  profile_max_steps: 200

optim:
  lr: 8e-5
  weight_decay: 1e-4
  betas: [0.9, 0.999]

scheduler:
  type: "cosine"
  warmup_steps: 500
  min_lr: 0.05

ema:
  enable: true
  decay: 0.9999

loss:
  l1_weight: 1.0
  charbonnier_weight: 0.5
  ssim_weight: 0.1
  flow_smooth_weight: 0.01
  flow_edge_weight: 0.02
  flow_edge_alpha: 10.0
  warp_weight: 0.01
  occ_reg_weight: 0.01
  grad_weight: 0.01
  laplacian_weight: 0.0
  perceptual_weight: 0.03
  perceptual_type: "lpips"
  lpips_net: "vgg"
  perceptual_layers: [3, 8, 15]
  multiscale_weight: 0.2
  multiscale_scales: [0.5, 0.25]
  multiscale_weights: [0.5, 0.25]
  temporal_weight: 0.12
  temporal_t: 0.25
  temporal_every: 4
  temporal_detach: true
  charbonnier_eps: 0.001
  color_consistency_weight: 0.02
  brightness_consistency_weight: 0.02
  color_consistency_use_std: true
  brightness_consistency_use_std: true

model:
  base_channels: 64 # 64
  num_scales: 4 # 4
  flow_refine_depth: 3 # 3
  fusion_depth: 3 # 3
  use_attention: true
  attention_heads: 4
  attention_window: 8
  use_occlusion_mask: true

infer:
  pad_multiple: 8