default.yml 2.3 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115
  1. data:
  2. batch_size: 32
  3. pin_memory: true
  4. num_workers: 6
  5. # Thomas said it should be at least about 5-10x your batch size; beyond that,
  6. # the differences become academic.
  7. shuffle_buffer: 1250
  8. seed: ${train.seed}
  9. dataset:
  10. meta:
  11. gcc3m:
  12. type: img_txt_pair
  13. path: local_data/gcc3m_shards
  14. prefix: gcc-train-{000000..00436}.tar
  15. length: 2891445
  16. gcc12m:
  17. type: img_txt_pair
  18. path: local_data/gcc12m_shards
  19. prefix: gcc-conceptual-12m-{000000..001943}.tar
  20. length: 11156203
  21. yfcc14m:
  22. type: img_txt_pair
  23. path: local_data/yfcc14m_shards
  24. prefix: yfcc14m-{000000..001888}.tar
  25. length: 14615499
  26. redcap12m:
  27. type: img_txt_pair
  28. path: local_data/redcap12m_shards
  29. prefix: redcap12m-{000000..001211}.tar
  30. length: 11866987
  31. imagenet:
  32. type: img_cls_pair
  33. path: local_data/imagenet_shards
  34. prefix: imagenet-val-{000000..000049}.tar
  35. length: 50000
  36. cuhkpedes:
  37. type: img_txt_pair
  38. path: local_data/cuhkpedes_shards
  39. prefix: cuhkpedes-train-{000000..000255}.tar
  40. length: 34054
  41. train:
  42. # - gcc3m
  43. - gcc12m
  44. - yfcc14m
  45. - cuhkpedes
  46. val:
  47. - imagenet
  48. img_aug:
  49. deit_aug: true
  50. img_size: 224
  51. img_scale: [0.08, 1.0]
  52. interpolation: bilinear
  53. color_jitter: 0.4
  54. auto_augment: 'rand-m9-mstd0.5-inc1'
  55. re_prob: 0.25
  56. re_mode: 'pixel'
  57. re_count: 1
  58. text_aug:
  59. max_seq_len: 77
  60. multi_label: 0
  61. word_type: 'noun'
  62. train:
  63. start_epoch: 0
  64. epochs: 30
  65. warmup_epochs: 2
  66. base_lr: 1.6e-3
  67. weight_decay: 0.05
  68. warmup_lr: 4e-6
  69. min_lr: 4e-5
  70. clip_grad: 5.0
  71. accumulation_steps: 0
  72. amp_opt_level: O1
  73. seed: 0
  74. lr_scheduler:
  75. name: cosine
  76. optimizer:
  77. name: adamw
  78. eps: 1e-8
  79. betas: [0.9, 0.999]
  80. evaluate:
  81. eval_only: false
  82. eval_freq: 1
  83. task:
  84. - cls
  85. - seg
  86. cls:
  87. save_best: true
  88. template: subset
  89. seg:
  90. save_best: true
  91. cfg: segmentation/configs/_base_/datasets/pascal_voc12.py
  92. template: simple
  93. opts: []
  94. checkpoint:
  95. auto_resume: true
  96. resume: ''
  97. freq: 1
  98. max_kept: -1
  99. save_freq: 1
  100. model_name: '' # display name in the logger
  101. output: ???
  102. tag: default
  103. print_freq: 10
  104. seed: 0
  105. wandb: false
  106. local_rank: ???
  107. vis: []