PP-OCRv4_mobile_rec.yaml 2.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138
  1. Global:
  2. debug: false
  3. use_gpu: true
  4. epoch_num: 20
  5. log_smooth_window: 20
  6. print_batch_step: 20
  7. save_model_dir: ./output/rec_ppocr_v4
  8. save_epoch_step: 100
  9. eval_batch_step: [0, 100]
  10. cal_metric_during_train: true
  11. pretrained_model: https://paddleocr.bj.bcebos.com/pretrained/ch_PP-OCRv4_rec_trained.pdparams
  12. checkpoints:
  13. save_inference_dir:
  14. use_visualdl: false
  15. infer_img: doc/imgs_words/ch/word_1.jpg
  16. character_dict_path: ppocr/utils/ppocr_keys_v1.txt
  17. max_text_length: &max_text_length 25
  18. infer_mode: false
  19. use_space_char: true
  20. distributed: true
  21. save_res_path: ./output/rec/predicts_ppocrv3.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: 0.001
  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: *max_text_length
  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. Train:
  68. dataset:
  69. name: MSTextRecDataset
  70. ds_width: false
  71. data_dir: ./train_data/
  72. ext_op_transform_idx: 1
  73. label_file_list:
  74. - ./train_data/train_list.txt
  75. transforms:
  76. - DecodeImage:
  77. img_mode: BGR
  78. channel_first: false
  79. - RecConAug:
  80. prob: 0.5
  81. ext_data_num: 2
  82. image_shape: [48, 320, 3]
  83. max_text_length: *max_text_length
  84. - RecAug:
  85. - MultiLabelEncode:
  86. gtc_encode: NRTRLabelEncode
  87. - KeepKeys:
  88. keep_keys:
  89. - image
  90. - label_ctc
  91. - label_gtc
  92. - length
  93. - valid_ratio
  94. sampler:
  95. name: MultiScaleSampler
  96. scales: [[320, 32], [320, 48], [320, 64]]
  97. first_bs: 8
  98. fix_bs: false
  99. divided_factor: [8, 16] # w, h
  100. is_training: True
  101. loader:
  102. shuffle: true
  103. batch_size_per_card: 8
  104. drop_last: true
  105. num_workers: 8
  106. Eval:
  107. dataset:
  108. name: TextRecDataset
  109. data_dir: ./train_data
  110. label_file_list:
  111. - ./train_data/val_list.txt
  112. transforms:
  113. - DecodeImage:
  114. img_mode: BGR
  115. channel_first: false
  116. - MultiLabelEncode:
  117. gtc_encode: NRTRLabelEncode
  118. - RecResizeImg:
  119. image_shape: [3, 48, 320]
  120. - KeepKeys:
  121. keep_keys:
  122. - image
  123. - label_ctc
  124. - label_gtc
  125. - length
  126. - valid_ratio
  127. loader:
  128. shuffle: false
  129. drop_last: false
  130. batch_size_per_card: 128
  131. num_workers: 4