PP-YOLOE_plus_SOD-S.yaml 3.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164
  1. # Runtime
  2. use_gpu: true
  3. use_xpu: false
  4. use_mlu: false
  5. use_npu: false
  6. save_dir: output
  7. print_flops: false
  8. print_params: false
  9. log_iter: 100
  10. snapshot_epoch: 10
  11. use_ema: true
  12. # Dataset
  13. metric: COCO
  14. num_classes: 10
  15. TrainDataset:
  16. name: COCODataSet
  17. image_dir: VisDrone2019-DET-train
  18. anno_path: train.json
  19. dataset_dir: dataset/visdrone
  20. data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
  21. EvalDataset:
  22. name: COCODataSet
  23. image_dir: VisDrone2019-DET-val
  24. anno_path: val.json
  25. dataset_dir: dataset/visdrone
  26. TestDataset:
  27. name: ImageFolder
  28. anno_path: val.json
  29. dataset_dir: dataset/visdrone
  30. #reader
  31. worker_num: 4
  32. eval_height: &eval_height 640
  33. eval_width: &eval_width 640
  34. eval_size: &eval_size [*eval_height, *eval_width]
  35. TrainReader:
  36. sample_transforms:
  37. - Decode: {}
  38. - RandomDistort: {}
  39. - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
  40. - RandomCrop: {}
  41. - RandomFlip: {}
  42. batch_transforms:
  43. - BatchRandomResize: {target_size: [320, 352, 384, 416, 448, 480, 512, 544, 576, 608, 640, 672, 704, 736, 768], random_size: True, random_interp: True, keep_ratio: False}
  44. - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
  45. - Permute: {}
  46. - PadGT: {}
  47. batch_size: 8
  48. shuffle: true
  49. drop_last: true
  50. use_shared_memory: true
  51. collate_batch: true
  52. EvalReader:
  53. sample_transforms:
  54. - Decode: {}
  55. - Resize: {target_size: *eval_size, keep_ratio: False, interp: 2}
  56. - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
  57. - Permute: {}
  58. batch_size: 1
  59. TestReader:
  60. inputs_def:
  61. image_shape: [3, *eval_height, *eval_width]
  62. sample_transforms:
  63. - Decode: {}
  64. - Resize: {target_size: *eval_size, keep_ratio: False, interp: 2}
  65. - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
  66. - Permute: {}
  67. batch_size: 1
  68. fuse_normalize: True
  69. # Model
  70. pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_s_80e_coco.pdparams
  71. depth_mult: 0.33
  72. width_mult: 0.50
  73. architecture: YOLOv3
  74. norm_type: sync_bn
  75. use_ema: true
  76. ema_decay: 0.9998
  77. ema_black_list: ['proj_conv.weight']
  78. custom_black_list: ['reduce_mean']
  79. YOLOv3:
  80. backbone: CSPResNet
  81. neck: CustomCSPPAN
  82. yolo_head: PPYOLOEHead
  83. post_process: ~
  84. CSPResNet:
  85. layers: [3, 6, 6, 3]
  86. channels: [64, 128, 256, 512, 1024]
  87. return_idx: [1, 2, 3]
  88. use_large_stem: True
  89. use_alpha: True
  90. CustomCSPPAN:
  91. out_channels: [768, 384, 192]
  92. stage_num: 1
  93. block_num: 3
  94. act: 'swish'
  95. spp: true
  96. num_layers: 4
  97. use_trans: True
  98. PPYOLOEHead:
  99. reg_range: [-2,8]
  100. static_assigner_epoch: -1
  101. fpn_strides: [32, 16, 8]
  102. grid_cell_scale: 5.0
  103. grid_cell_offset: 0.5
  104. use_varifocal_loss: True
  105. loss_weight: {class: 1.0, iou: 2.5, dfl: 0.5}
  106. static_assigner:
  107. name: ATSSAssigner
  108. topk: 9
  109. assigner:
  110. name: TaskAlignedAssigner_CR
  111. center_radius: 1
  112. topk: 13
  113. alpha: 1.0
  114. beta: 6.0
  115. nms:
  116. name: MultiClassNMS
  117. nms_top_k: 10000
  118. keep_top_k: 500
  119. score_threshold: 0.01
  120. nms_threshold: 0.6
  121. # Optimizer
  122. epoch: 80
  123. LearningRate:
  124. base_lr: 0.01
  125. schedulers:
  126. - !CosineDecay
  127. max_epochs: 96
  128. - !LinearWarmup
  129. start_factor: 0.
  130. epochs: 1
  131. OptimizerBuilder:
  132. optimizer:
  133. momentum: 0.9
  134. type: Momentum
  135. regularizer:
  136. factor: 0.0005
  137. type: L2
  138. # Exporting the model
  139. export:
  140. post_process: True # Whether post-processing is included in the network when export model.
  141. nms: True # Whether NMS is included in the network when export model.
  142. benchmark: False # It is used to testing model performance, if set `True`, post-process and NMS will not be exported.
  143. fuse_conv_bn: False