SLANeXt_wired.yaml 3.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179
  1. Global:
  2. use_gpu: true
  3. epoch_num: 400
  4. log_smooth_window: 20
  5. print_batch_step: 20
  6. save_model_dir: ./output/SLANeXt_wired
  7. save_epoch_step: 400
  8. eval_batch_step:
  9. - 0
  10. - 331
  11. cal_metric_during_train: true
  12. pretrained_model: null
  13. checkpoints: null
  14. save_inference_dir: ./output/SLANeXt_wired/infer
  15. use_visualdl: false
  16. infer_img: ppstructure/docs/table/table.jpg
  17. character_dict_path: ppocr/utils/dict/table_structure_dict_ch.txt
  18. character_type: en
  19. max_text_length: 500
  20. box_format: xyxyxyxy
  21. infer_mode: false
  22. use_sync_bn: true
  23. save_res_path: output/infer
  24. Optimizer:
  25. name: AdamW
  26. beta1: 0.9
  27. beta2: 0.999
  28. clip_norm: 5.0
  29. lr:
  30. name: Cosine
  31. learning_rate: 0.0001
  32. warmup_epoch: 1
  33. regularizer:
  34. name: L2
  35. factor: 0.0
  36. Architecture:
  37. model_type: table
  38. algorithm: SLANeXt
  39. Backbone:
  40. name: Vary_VIT_B
  41. image_size: 512
  42. encoder_embed_dim: 768
  43. encoder_depth: 12
  44. encoder_num_heads: 12
  45. encoder_global_attn_indexes: [2, 5, 8, 11]
  46. Head:
  47. name: SLAHead
  48. hidden_size: 512
  49. max_text_length: 500
  50. loc_reg_num: 8
  51. Loss:
  52. name: SLALoss
  53. structure_weight: 1.0
  54. # SLANeXt does not train the cell location task by default, set the loc_weight if needed.
  55. loc_weight: 0.0
  56. loc_loss: smooth_l1
  57. PostProcess:
  58. name: TableLabelDecode
  59. merge_no_span_structure: true
  60. Metric:
  61. name: TableMetric
  62. main_indicator: acc
  63. compute_bbox_metric: false
  64. loc_reg_num: 8
  65. box_format: xyxyxyxy
  66. del_thead_tbody: true
  67. Train:
  68. dataset:
  69. name: PubTabDataSet
  70. data_dir: train_data/table/train/
  71. label_file_list:
  72. - train_data/table/train.txt
  73. ratio_list:
  74. - 1
  75. transforms:
  76. - DecodeImage:
  77. img_mode: BGR
  78. channel_first: false
  79. - TableLabelEncode:
  80. learn_empty_box: false
  81. merge_no_span_structure: true
  82. replace_empty_cell_token: false
  83. loc_reg_num: 8
  84. max_text_length: 500
  85. - TableBoxEncode:
  86. in_box_format: xyxyxyxy
  87. out_box_format: xyxyxyxy
  88. - ResizeTableImage:
  89. max_len: 512
  90. resize_bboxes: true
  91. - NormalizeImage:
  92. scale: 1./255.
  93. mean:
  94. - 0.485
  95. - 0.456
  96. - 0.406
  97. std:
  98. - 0.229
  99. - 0.224
  100. - 0.225
  101. order: hwc
  102. - PaddingTableImage:
  103. size:
  104. - 512
  105. - 512
  106. - ToCHWImage: null
  107. - KeepKeys:
  108. keep_keys:
  109. - image
  110. - structure
  111. - bboxes
  112. - bbox_masks
  113. - length
  114. - shape
  115. loader:
  116. shuffle: true
  117. batch_size_per_card: 48
  118. drop_last: true
  119. num_workers: 1
  120. Eval:
  121. dataset:
  122. name: PubTabDataSet
  123. data_dir: train_data/table/val/
  124. label_file_list:
  125. - train_data/table/val.txt
  126. transforms:
  127. - DecodeImage:
  128. img_mode: BGR
  129. channel_first: false
  130. - TableLabelEncode:
  131. learn_empty_box: false
  132. merge_no_span_structure: true
  133. replace_empty_cell_token: false
  134. loc_reg_num: 8
  135. max_text_length: 500
  136. - TableBoxEncode:
  137. in_box_format: xyxyxyxy
  138. out_box_format: xyxyxyxy
  139. - ResizeTableImage:
  140. max_len: 512
  141. resize_bboxes: true
  142. - NormalizeImage:
  143. scale: 1./255.
  144. mean:
  145. - 0.485
  146. - 0.456
  147. - 0.406
  148. std:
  149. - 0.229
  150. - 0.224
  151. - 0.225
  152. order: hwc
  153. - PaddingTableImage:
  154. size:
  155. - 512
  156. - 512
  157. - ToCHWImage: null
  158. - KeepKeys:
  159. keep_keys:
  160. - image
  161. - structure
  162. - bboxes
  163. - bbox_masks
  164. - length
  165. - shape
  166. loader:
  167. shuffle: false
  168. drop_last: false
  169. batch_size_per_card: 48
  170. num_workers: 1
  171. profiler_options: null