default.yml 2.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131
  1. data:
  2. batch_size: 256
  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: 10000
  8. seed: ${train.seed}
  9. bpe_path: /home/linkslinks/文档/ai/GroupViT/datasets/bpe_simple_vocab_16e6.txt.gz
  10. dataset:
  11. meta:
  12. gcc3m:
  13. type: img_txt_pair
  14. path: local_data/gcc3m_shards
  15. prefix: gcc-train-{000000..00436}.tar
  16. length: 2891445
  17. gcc12m:
  18. type: img_txt_pair
  19. path: local_data/gcc12m_shards
  20. prefix: gcc-conceptual-12m-{000000..001943}.tar
  21. length: 11156203
  22. yfcc14m:
  23. type: img_txt_pair
  24. path: local_data/yfcc14m_shards
  25. prefix: yfcc14m-{000000..001888}.tar
  26. length: 14615499
  27. redcap12m:
  28. type: img_txt_pair
  29. path: local_data/redcap12m_shards
  30. prefix: redcap12m-{000000..001211}.tar
  31. length: 11866987
  32. imagenet:
  33. type: img_cls_pair
  34. path: local_data/imagenet_shards
  35. prefix: imagenet-val-{000000..000049}.tar
  36. length: 50000
  37. cuhkpedes_train:
  38. name: CUHK-PEDES
  39. type: img_txt_pair
  40. path: local_data/cuhkpedes_shards
  41. prefix: cuhkpedes-train-{000000..000004}.tar
  42. length: 34054
  43. cuhkpedes_val:
  44. raw_path: /home/linkslinks/dataset/
  45. name: CUHK-PEDES
  46. type: img_txt_pair
  47. path: local_data/cuhkpedes_shards
  48. prefix: cuhkpedes-val-{000000..000000}.tar
  49. length: 3078
  50. train:
  51. # - gcc3m
  52. # - gcc12m
  53. # - yfcc14m
  54. - cuhkpedes_train
  55. val:
  56. # - imagenet
  57. - cuhkpedes_val
  58. img_aug:
  59. deit_aug: true
  60. img_size: 224
  61. img_scale: [0.08, 1.0]
  62. interpolation: bilinear
  63. color_jitter: 0.4
  64. auto_augment: 'rand-m9-mstd0.5-inc1'
  65. re_prob: 0.25
  66. re_mode: 'pixel'
  67. re_count: 1
  68. text_aug:
  69. max_seq_len: 77
  70. multi_label: 0
  71. word_type: 'noun'
  72. train:
  73. start_epoch: 0
  74. epochs: 30
  75. warmup_epochs: 2
  76. base_lr: 1.6e-3
  77. weight_decay: 0.05
  78. warmup_lr: 4e-6
  79. min_lr: 4e-5
  80. clip_grad: 5.0
  81. accumulation_steps: 0
  82. amp_opt_level: O2
  83. seed: 0
  84. lr_scheduler:
  85. name: cosine
  86. optimizer:
  87. name: adamw
  88. eps: 1e-8
  89. betas: [0.9, 0.999]
  90. evaluate:
  91. eval_only: false
  92. eval_freq: 1
  93. task:
  94. - cls
  95. - seg
  96. - retrieval
  97. cls:
  98. save_best: true
  99. template: subset
  100. seg:
  101. save_best: true
  102. cfg: segmentation/configs/_base_/datasets/pascal_voc12.py
  103. template: simple
  104. opts: []
  105. retrieval:
  106. save_best: true
  107. template: simple
  108. opts: []
  109. checkpoint:
  110. auto_resume: true
  111. resume: ''
  112. freq: 1
  113. max_kept: -1
  114. save_freq: 1
  115. model_name: '' # display name in the logger
  116. output: ???
  117. tag: default
  118. print_freq: 10
  119. seed: 0
  120. wandb: false
  121. local_rank: ???
  122. vis: []