SLANeXt_wired.yaml 3.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180
  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. d2s_train_image_shape: [3, 512, 512]
  25. Optimizer:
  26. name: AdamW
  27. beta1: 0.9
  28. beta2: 0.999
  29. clip_norm: 5.0
  30. lr:
  31. name: Cosine
  32. learning_rate: 0.0001
  33. warmup_epoch: 1
  34. regularizer:
  35. name: L2
  36. factor: 0.0
  37. Architecture:
  38. model_type: table
  39. algorithm: SLANeXt
  40. Backbone:
  41. name: Vary_VIT_B
  42. image_size: 512
  43. encoder_embed_dim: 768
  44. encoder_depth: 12
  45. encoder_num_heads: 12
  46. encoder_global_attn_indexes: [2, 5, 8, 11]
  47. Head:
  48. name: SLAHead
  49. hidden_size: 512
  50. max_text_length: 500
  51. loc_reg_num: 8
  52. Loss:
  53. name: SLALoss
  54. structure_weight: 1.0
  55. # SLANeXt does not train the cell location task by default, set the loc_weight if needed.
  56. loc_weight: 0.0
  57. loc_loss: smooth_l1
  58. PostProcess:
  59. name: TableLabelDecode
  60. merge_no_span_structure: true
  61. Metric:
  62. name: TableMetric
  63. main_indicator: acc
  64. compute_bbox_metric: false
  65. loc_reg_num: 8
  66. box_format: xyxyxyxy
  67. del_thead_tbody: true
  68. Train:
  69. dataset:
  70. name: PubTabDataSet
  71. data_dir: train_data/table/train/
  72. label_file_list:
  73. - train_data/table/train.txt
  74. ratio_list:
  75. - 1
  76. transforms:
  77. - DecodeImage:
  78. img_mode: BGR
  79. channel_first: false
  80. - TableLabelEncode:
  81. learn_empty_box: false
  82. merge_no_span_structure: true
  83. replace_empty_cell_token: false
  84. loc_reg_num: 8
  85. max_text_length: 500
  86. - TableBoxEncode:
  87. in_box_format: xyxyxyxy
  88. out_box_format: xyxyxyxy
  89. - ResizeTableImage:
  90. max_len: 512
  91. resize_bboxes: True
  92. - NormalizeImage:
  93. scale: 1./255.
  94. mean:
  95. - 0.485
  96. - 0.456
  97. - 0.406
  98. std:
  99. - 0.229
  100. - 0.224
  101. - 0.225
  102. order: hwc
  103. - PaddingTableImage:
  104. size:
  105. - 512
  106. - 512
  107. - ToCHWImage: null
  108. - KeepKeys:
  109. keep_keys:
  110. - image
  111. - structure
  112. - bboxes
  113. - bbox_masks
  114. - length
  115. - shape
  116. loader:
  117. shuffle: true
  118. batch_size_per_card: 48
  119. drop_last: true
  120. num_workers: 1
  121. Eval:
  122. dataset:
  123. name: PubTabDataSet
  124. data_dir: train_data/table/val/
  125. label_file_list:
  126. - train_data/table/val.txt
  127. transforms:
  128. - DecodeImage:
  129. img_mode: BGR
  130. channel_first: false
  131. - TableLabelEncode:
  132. learn_empty_box: false
  133. merge_no_span_structure: true
  134. replace_empty_cell_token: false
  135. loc_reg_num: 8
  136. max_text_length: 500
  137. - TableBoxEncode:
  138. in_box_format: xyxyxyxy
  139. out_box_format: xyxyxyxy
  140. - ResizeTableImage:
  141. max_len: 512
  142. resize_bboxes: True
  143. - NormalizeImage:
  144. scale: 1./255.
  145. mean:
  146. - 0.485
  147. - 0.456
  148. - 0.406
  149. std:
  150. - 0.229
  151. - 0.224
  152. - 0.225
  153. order: hwc
  154. - PaddingTableImage:
  155. size:
  156. - 512
  157. - 512
  158. - ToCHWImage: null
  159. - KeepKeys:
  160. keep_keys:
  161. - image
  162. - structure
  163. - bboxes
  164. - bbox_masks
  165. - length
  166. - shape
  167. loader:
  168. shuffle: false
  169. drop_last: false
  170. batch_size_per_card: 48
  171. num_workers: 1
  172. profiler_options: null