en_PP-OCRv4_mobile_rec.yaml 2.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150
  1. Global:
  2. debug: false
  3. use_gpu: true
  4. epoch_num: 50
  5. log_smooth_window: 20
  6. print_batch_step: 10
  7. save_model_dir: ./output/rec_ppocr_v4
  8. save_epoch_step: 10
  9. eval_batch_step:
  10. - 0
  11. - 2000
  12. cal_metric_during_train: true
  13. pretrained_model: null
  14. checkpoints: null
  15. save_inference_dir: null
  16. use_visualdl: false
  17. infer_img: doc/imgs_words/ch/word_1.jpg
  18. character_dict_path: ppocr/utils/en_dict.txt
  19. max_text_length: 25
  20. infer_mode: false
  21. use_space_char: true
  22. distributed: true
  23. save_res_path: ./output/rec/predicts_ppocrv3.txt
  24. Optimizer:
  25. name: Adam
  26. beta1: 0.9
  27. beta2: 0.999
  28. lr:
  29. name: Cosine
  30. learning_rate: 0.0005
  31. warmup_epoch: 5
  32. regularizer:
  33. name: L2
  34. factor: 3.0e-05
  35. Architecture:
  36. model_type: rec
  37. algorithm: SVTR_LCNet
  38. Transform: null
  39. Backbone:
  40. name: PPLCNetV3
  41. scale: 0.95
  42. Head:
  43. name: MultiHead
  44. head_list:
  45. - CTCHead:
  46. Neck:
  47. name: svtr
  48. dims: 120
  49. depth: 2
  50. hidden_dims: 120
  51. kernel_size:
  52. - 1
  53. - 3
  54. use_guide: true
  55. Head:
  56. fc_decay: 1.0e-05
  57. - NRTRHead:
  58. nrtr_dim: 384
  59. max_text_length: 25
  60. Loss:
  61. name: MultiLoss
  62. loss_config_list:
  63. - CTCLoss: null
  64. - NRTRLoss: null
  65. PostProcess:
  66. name: CTCLabelDecode
  67. Metric:
  68. name: RecMetric
  69. main_indicator: acc
  70. ignore_space: false
  71. Train:
  72. dataset:
  73. name: MultiScaleDataSet
  74. ds_width: false
  75. data_dir: ./train_data/
  76. ext_op_transform_idx: 1
  77. label_file_list:
  78. - ./train_data/train_list.txt
  79. transforms:
  80. - DecodeImage:
  81. img_mode: BGR
  82. channel_first: false
  83. - RecConAug:
  84. prob: 0.5
  85. ext_data_num: 2
  86. image_shape:
  87. - 48
  88. - 320
  89. - 3
  90. max_text_length: 25
  91. - RecAug: null
  92. - MultiLabelEncode:
  93. gtc_encode: NRTRLabelEncode
  94. - KeepKeys:
  95. keep_keys:
  96. - image
  97. - label_ctc
  98. - label_gtc
  99. - length
  100. - valid_ratio
  101. sampler:
  102. name: MultiScaleSampler
  103. scales:
  104. - - 320
  105. - 32
  106. - - 320
  107. - 48
  108. - - 320
  109. - 64
  110. first_bs: 96
  111. fix_bs: false
  112. divided_factor:
  113. - 8
  114. - 16
  115. is_training: true
  116. loader:
  117. shuffle: true
  118. batch_size_per_card: 96
  119. drop_last: true
  120. num_workers: 8
  121. Eval:
  122. dataset:
  123. name: SimpleDataSet
  124. data_dir: ./train_data
  125. label_file_list:
  126. - ./train_data/val_list.txt
  127. transforms:
  128. - DecodeImage:
  129. img_mode: BGR
  130. channel_first: false
  131. - MultiLabelEncode:
  132. gtc_encode: NRTRLabelEncode
  133. - RecResizeImg:
  134. image_shape:
  135. - 3
  136. - 48
  137. - 320
  138. - KeepKeys:
  139. keep_keys:
  140. - image
  141. - label_ctc
  142. - label_gtc
  143. - length
  144. - valid_ratio
  145. loader:
  146. shuffle: false
  147. drop_last: false
  148. batch_size_per_card: 128
  149. num_workers: 4
  150. profiler_options: null