data:
  batch_size: 256
  pin_memory: true
  num_workers: 6
  # Thomas said it should be at least about 5-10x your batch size; beyond that,
  # the differences become academic.
  shuffle_buffer: 10000
  seed: ${train.seed}
  dataset:
    meta:
      gcc3m:
        type: img_txt_pair
        path: local_data/gcc3m_shards
        prefix: gcc-train-{000000..00436}.tar
        length: 2891445
      gcc12m:
        type: img_txt_pair
        path: local_data/gcc12m_shards
        prefix: gcc-conceptual-12m-{000000..001943}.tar
        length: 11156203
      yfcc14m:
        type: img_txt_pair
        path: local_data/yfcc14m_shards
        prefix: yfcc14m-{000000..001888}.tar
        length: 14615499
      redcap12m:
        type: img_txt_pair
        path: local_data/redcap12m_shards
        prefix: redcap12m-{000000..001211}.tar
        length: 11866987
      imagenet:
        type: img_cls_pair
        path: local_data/imagenet_shards
        prefix: imagenet-val-{000000..000049}.tar
        length: 50000
      cuhkpedes_train:
        type: img_txt_pair
        path: local_data/cuhkpedes_shards
        prefix: cuhkpedes-train-{000000..000255}.tar
        length: 34054
      cuhkpedes_val:
        type: img_txt_pair
        path: local_data/cuhkpedes_shards
        prefix: cuhkpedes-val-{000000..000023}.tar
        length: 3078
    train:
      # - gcc3m
      # - gcc12m
      # - yfcc14m
      - cuhkpedes_train
    val:
      # - imagenet
      - cuhkpedes_val

  img_aug:
    deit_aug: true
    img_size: 224
    img_scale: [0.08, 1.0]
    interpolation: bilinear
    color_jitter: 0.4
    auto_augment: 'rand-m9-mstd0.5-inc1'
    re_prob: 0.25
    re_mode: 'pixel'
    re_count: 1
  text_aug:
    max_seq_len: 77
    multi_label: 0
    word_type: 'noun'

train:
  start_epoch: 0
  epochs: 30
  warmup_epochs: 2
  base_lr: 1.6e-3
  weight_decay: 0.05
  warmup_lr: 4e-6
  min_lr: 4e-5
  clip_grad: 5.0
  accumulation_steps: 0
  amp_opt_level: O2
  seed: 0

  lr_scheduler:
    name: cosine

  optimizer:
    name: adamw
    eps: 1e-8
    betas: [0.9, 0.999]

evaluate:
  eval_only: false
  eval_freq: 1
  task:
    - cls
    - seg
  cls:
    save_best: true
    template: subset
  seg:
    save_best: true
    cfg: segmentation/configs/_base_/datasets/pascal_voc12.py
    template: simple
    opts: []

checkpoint:
  auto_resume: true
  resume: ''
  freq: 1
  max_kept: -1
  save_freq: 1


model_name: '' # display name in the logger
output: ???
tag: default
print_freq: 10
seed: 0
wandb: false
local_rank: ???
vis: []