PP-FormulaNet-L.yaml 2.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117
  1. Global:
  2. use_gpu: True
  3. epoch_num: 10
  4. log_smooth_window: 10
  5. print_batch_step: 10
  6. save_model_dir: ./output/rec/pp_formulanet_l/
  7. save_epoch_step: 2
  8. # evaluation is run every 417 iterations (1 epoch)(batch_size = 24) # max_seq_len: 1024
  9. eval_batch_step: [0, 417 ]
  10. cal_metric_during_train: True
  11. pretrained_model:
  12. checkpoints:
  13. save_inference_dir:
  14. use_visualdl: False
  15. infer_img: doc/datasets/pme_demo/0000013.png
  16. infer_mode: False
  17. use_space_char: False
  18. rec_char_dict_path: &rec_char_dict_path ppocr/utils/dict/unimernet_tokenizer
  19. max_new_tokens: &max_new_tokens 1024
  20. input_size: &input_size [768, 768]
  21. save_res_path: ./output/rec/predicts_unimernet_latexocr.txt
  22. allow_resize_largeImg: False
  23. start_ema: True
  24. Optimizer:
  25. name: AdamW
  26. beta1: 0.9
  27. beta2: 0.999
  28. weight_decay: 0.05
  29. lr:
  30. name: LinearWarmupCosine
  31. learning_rate: 0.0001
  32. Architecture:
  33. model_type: rec
  34. algorithm: PP-FormulaNet-L
  35. in_channels: 3
  36. Transform:
  37. Backbone:
  38. name: Vary_VIT_B_Formula
  39. image_size: 768
  40. encoder_embed_dim: 768
  41. encoder_depth: 12
  42. encoder_num_heads: 12
  43. encoder_global_attn_indexes: [2, 5, 8, 11]
  44. Head:
  45. name: PPFormulaNet_Head
  46. max_new_tokens: *max_new_tokens
  47. decoder_start_token_id: 0
  48. decoder_ffn_dim: 2048
  49. decoder_hidden_size: 512
  50. decoder_layers: 8
  51. temperature: 0.2
  52. do_sample: False
  53. top_p: 0.95
  54. encoder_hidden_size: 1024
  55. is_export: False
  56. length_aware: False
  57. use_parallel: False
  58. parallel_step: 0
  59. Loss:
  60. name: PPFormulaNet_L_Loss
  61. PostProcess:
  62. name: UniMERNetDecode
  63. rec_char_dict_path: *rec_char_dict_path
  64. Metric:
  65. name: LaTeXOCRMetric
  66. main_indicator: exp_rate
  67. cal_blue_score: False
  68. Train:
  69. dataset:
  70. name: SimpleDataSet
  71. data_dir: ./ocr_rec_latexocr_dataset_example
  72. label_file_list: ["./ocr_rec_latexocr_dataset_example/train.txt"]
  73. transforms:
  74. - UniMERNetImgDecode:
  75. input_size: *input_size
  76. - UniMERNetTrainTransform:
  77. - LatexImageFormat:
  78. - UniMERNetLabelEncode:
  79. rec_char_dict_path: *rec_char_dict_path
  80. max_seq_len: *max_new_tokens
  81. - KeepKeys:
  82. keep_keys: ['image', 'label', 'attention_mask']
  83. loader:
  84. shuffle: False
  85. drop_last: False
  86. batch_size_per_card: 6
  87. num_workers: 0
  88. collate_fn: UniMERNetCollator
  89. Eval:
  90. dataset:
  91. name: SimpleDataSet
  92. data_dir: ./ocr_rec_latexocr_dataset_example
  93. label_file_list: ["./ocr_rec_latexocr_dataset_example/val.txt"]
  94. transforms:
  95. - UniMERNetImgDecode:
  96. input_size: *input_size
  97. - UniMERNetTestTransform:
  98. - LatexImageFormat:
  99. - UniMERNetLabelEncode:
  100. max_seq_len: *max_new_tokens
  101. rec_char_dict_path: *rec_char_dict_path
  102. - KeepKeys:
  103. keep_keys: ['image', 'label', 'attention_mask', 'filename']
  104. loader:
  105. shuffle: False
  106. drop_last: False
  107. batch_size_per_card: 10
  108. num_workers: 0
  109. collate_fn: UniMERNetCollator