ch_RepSVTR_rec.yaml 2.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135
  1. Global:
  2. debug: false
  3. use_gpu: true
  4. epoch_num: 200
  5. log_smooth_window: 20
  6. print_batch_step: 10
  7. save_model_dir: ./output/rec_repsvtr_ch
  8. save_epoch_step: 10
  9. eval_batch_step: [0, 1000]
  10. cal_metric_during_train: False
  11. pretrained_model: https://paddleocr.bj.bcebos.com/pretrained/ch_SVTRv2_rec_mobile_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_repsvtr.txt
  22. d2s_train_image_shape: [3,48,320]
  23. Optimizer:
  24. name: AdamW
  25. beta1: 0.9
  26. beta2: 0.999
  27. epsilon: 1.e-8
  28. weight_decay: 0.025
  29. no_weight_decay_name: norm
  30. one_dim_param_no_weight_decay: True
  31. lr:
  32. name: Cosine
  33. learning_rate: 0.001 # 8gpus 192bs
  34. warmup_epoch: 5
  35. Architecture:
  36. model_type: rec
  37. algorithm: SVTR_HGNet
  38. Transform:
  39. Backbone:
  40. name: RepSVTR
  41. Head:
  42. name: MultiHead
  43. head_list:
  44. - CTCHead:
  45. Neck:
  46. name: svtr
  47. dims: 256
  48. depth: 2
  49. hidden_dims: 256
  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. num_decoder_layers: 2
  58. Loss:
  59. name: MultiLoss
  60. loss_config_list:
  61. - CTCLoss:
  62. - NRTRLoss:
  63. PostProcess:
  64. name: CTCLabelDecode
  65. Metric:
  66. name: RecMetric
  67. main_indicator: acc
  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. - RecAug:
  81. - MultiLabelEncode:
  82. gtc_encode: NRTRLabelEncode
  83. - KeepKeys:
  84. keep_keys:
  85. - image
  86. - label_ctc
  87. - label_gtc
  88. - length
  89. - valid_ratio
  90. sampler:
  91. name: MultiScaleSampler
  92. scales: [[320, 32], [320, 48], [320, 64]]
  93. first_bs: &bs 192
  94. fix_bs: false
  95. divided_factor: [8, 16] # w, h
  96. is_training: True
  97. loader:
  98. shuffle: true
  99. batch_size_per_card: *bs
  100. drop_last: true
  101. num_workers: 8
  102. Eval:
  103. dataset:
  104. name: SimpleDataSet
  105. data_dir: ./train_data
  106. label_file_list:
  107. - ./train_data/val_list.txt
  108. transforms:
  109. - DecodeImage:
  110. img_mode: BGR
  111. channel_first: false
  112. - MultiLabelEncode:
  113. gtc_encode: NRTRLabelEncode
  114. - RecResizeImg:
  115. image_shape: [3, 48, 320]
  116. - KeepKeys:
  117. keep_keys:
  118. - image
  119. - label_ctc
  120. - label_gtc
  121. - length
  122. - valid_ratio
  123. loader:
  124. shuffle: false
  125. drop_last: false
  126. batch_size_per_card: 128
  127. num_workers: 4