PP-OCRv3_mobile_det.yaml 3.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162
  1. Global:
  2. debug: false
  3. use_gpu: true
  4. epoch_num: 500
  5. log_smooth_window: 20
  6. print_batch_step: 10
  7. save_model_dir: ./output/ch_PP-OCR_V3_det/
  8. save_epoch_step: 100
  9. eval_batch_step:
  10. - 0
  11. - 400
  12. cal_metric_during_train: false
  13. pretrained_model: ch_PP-OCRv3_det_distill_train/student.pdparams
  14. checkpoints: null
  15. save_inference_dir: null
  16. use_visualdl: false
  17. infer_img: doc/imgs_en/img_10.jpg
  18. save_res_path: ./checkpoints/det_db/predicts_db.txt
  19. distributed: true
  20. Architecture:
  21. model_type: det
  22. algorithm: DB
  23. Transform:
  24. Backbone:
  25. name: MobileNetV3
  26. scale: 0.5
  27. model_name: large
  28. disable_se: True
  29. Neck:
  30. name: RSEFPN
  31. out_channels: 96
  32. shortcut: True
  33. Head:
  34. name: DBHead
  35. k: 50
  36. Loss:
  37. name: DBLoss
  38. balance_loss: true
  39. main_loss_type: DiceLoss
  40. alpha: 5
  41. beta: 10
  42. ohem_ratio: 3
  43. Optimizer:
  44. name: Adam
  45. beta1: 0.9
  46. beta2: 0.999
  47. lr:
  48. name: Cosine
  49. learning_rate: 0.001
  50. warmup_epoch: 2
  51. regularizer:
  52. name: L2
  53. factor: 5.0e-05
  54. PostProcess:
  55. name: DBPostProcess
  56. thresh: 0.3
  57. box_thresh: 0.6
  58. max_candidates: 1000
  59. unclip_ratio: 1.5
  60. Metric:
  61. name: DetMetric
  62. main_indicator: hmean
  63. Train:
  64. dataset:
  65. name: TextDetDataset
  66. data_dir: datasets/ICDAR2015
  67. label_file_list:
  68. - datasets/ICDAR2015/train.txt
  69. transforms:
  70. - DecodeImage:
  71. img_mode: BGR
  72. channel_first: false
  73. - DetLabelEncode: null
  74. - IaaAugment:
  75. augmenter_args:
  76. - type: Fliplr
  77. args:
  78. p: 0.5
  79. - type: Affine
  80. args:
  81. rotate:
  82. - -10
  83. - 10
  84. - type: Resize
  85. args:
  86. size:
  87. - 0.5
  88. - 3
  89. - EastRandomCropData:
  90. size:
  91. - 960
  92. - 960
  93. max_tries: 50
  94. keep_ratio: true
  95. - MakeBorderMap:
  96. shrink_ratio: 0.4
  97. thresh_min: 0.3
  98. thresh_max: 0.7
  99. - MakeShrinkMap:
  100. shrink_ratio: 0.4
  101. min_text_size: 8
  102. - NormalizeImage:
  103. scale: 1./255.
  104. mean:
  105. - 0.485
  106. - 0.456
  107. - 0.406
  108. std:
  109. - 0.229
  110. - 0.224
  111. - 0.225
  112. order: hwc
  113. - ToCHWImage: null
  114. - KeepKeys:
  115. keep_keys:
  116. - image
  117. - threshold_map
  118. - threshold_mask
  119. - shrink_map
  120. - shrink_mask
  121. loader:
  122. shuffle: true
  123. drop_last: false
  124. batch_size_per_card: 8
  125. num_workers: 4
  126. Eval:
  127. dataset:
  128. name: TextDetDataset
  129. data_dir: datasets/ICDAR2015
  130. label_file_list:
  131. - datasets/ICDAR2015/val.txt
  132. transforms:
  133. - DecodeImage:
  134. img_mode: BGR
  135. channel_first: false
  136. - DetLabelEncode: null
  137. - DetResizeForTest: null
  138. - NormalizeImage:
  139. scale: 1./255.
  140. mean:
  141. - 0.485
  142. - 0.456
  143. - 0.406
  144. std:
  145. - 0.229
  146. - 0.224
  147. - 0.225
  148. order: hwc
  149. - ToCHWImage: null
  150. - KeepKeys:
  151. keep_keys:
  152. - image
  153. - shape
  154. - polys
  155. - ignore_tags
  156. loader:
  157. shuffle: false
  158. drop_last: false
  159. batch_size_per_card: 1
  160. num_workers: 2