latin_PP-OCRv5_mobile_rec.yaml 2.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143
  1. Global:
  2. model_name: latin_PP-OCRv5_mobile_rec # To use static model for inference.
  3. debug: false
  4. use_gpu: true
  5. epoch_num: 75
  6. log_smooth_window: 20
  7. print_batch_step: 10
  8. save_model_dir: ./output/latin_rec_ppocr_v5
  9. save_epoch_step: 10
  10. eval_batch_step: [0, 500]
  11. cal_metric_during_train: true
  12. pretrained_model:
  13. checkpoints:
  14. save_inference_dir:
  15. use_visualdl: false
  16. infer_img:
  17. character_dict_path: ppocr/utils/dict/ppocrv5_latin_dict.txt
  18. infer_mode: false
  19. use_space_char: true
  20. distributed: true
  21. save_res_path: ./output/rec/predicts_ppocrv5_latin.txt
  22. d2s_train_image_shape: [3, 48, 320]
  23. Optimizer:
  24. name: Adam
  25. beta1: 0.9
  26. beta2: 0.999
  27. lr:
  28. name: Cosine
  29. learning_rate: 0.0005
  30. warmup_epoch: 5
  31. regularizer:
  32. name: L2
  33. factor: 3.0e-05
  34. Architecture:
  35. model_type: rec
  36. algorithm: SVTR_LCNet
  37. Transform:
  38. Backbone:
  39. name: PPLCNetV3
  40. scale: 0.95
  41. Head:
  42. name: MultiHead
  43. head_list:
  44. - CTCHead:
  45. Neck:
  46. name: svtr
  47. dims: 120
  48. depth: 2
  49. hidden_dims: 120
  50. kernel_size: [1, 3]
  51. use_guide: True
  52. Head:
  53. fc_decay: 0.00001
  54. - NRTRHead:
  55. nrtr_dim: 384
  56. max_text_length: 25
  57. Loss:
  58. name: MultiLoss
  59. loss_config_list:
  60. - CTCLoss:
  61. - NRTRLoss:
  62. PostProcess:
  63. name: CTCLabelDecode
  64. Metric:
  65. name: RecMetric
  66. main_indicator: acc
  67. ignore_space: False
  68. Train:
  69. dataset:
  70. name: MultiScaleDataSet
  71. ds_width: false
  72. data_dir: ./train_data/
  73. ext_op_transform_idx: 1
  74. label_file_list:
  75. - ./train_data/train_list.txt
  76. transforms:
  77. - DecodeImage:
  78. img_mode: BGR
  79. channel_first: false
  80. - RecConAug:
  81. prob: 0.5
  82. ext_data_num: 2
  83. image_shape: [48, 320, 3]
  84. max_text_length: 25
  85. - RecAug:
  86. - MultiLabelEncode:
  87. max_text_length: 25
  88. gtc_encode: NRTRLabelEncode
  89. - KeepKeys:
  90. keep_keys:
  91. - image
  92. - label_ctc
  93. - label_gtc
  94. - length
  95. - valid_ratio
  96. sampler:
  97. name: MultiScaleSampler
  98. scales: [[320, 32], [320, 48], [320, 64]]
  99. first_bs: &bs 128
  100. fix_bs: false
  101. divided_factor: [8, 16] # w, h
  102. is_training: True
  103. loader:
  104. shuffle: true
  105. batch_size_per_card: *bs
  106. drop_last: true
  107. num_workers: 8
  108. Eval:
  109. dataset:
  110. name: SimpleDataSet
  111. data_dir: ./train_data/
  112. label_file_list:
  113. - ./train_data/eval_list.txt
  114. transforms:
  115. - DecodeImage:
  116. img_mode: BGR
  117. channel_first: false
  118. - MultiLabelEncode:
  119. max_text_length: 1000
  120. gtc_encode: NRTRLabelEncode
  121. - RecResizeImg:
  122. eval_mode: True
  123. image_shape: [3, 48, 320]
  124. - KeepKeys:
  125. keep_keys:
  126. - image
  127. - label_ctc
  128. - label_gtc
  129. - length
  130. - valid_ratio
  131. loader:
  132. shuffle: false
  133. drop_last: false
  134. batch_size_per_card: 1
  135. num_workers: 4