You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

reademe.md 2.4 kB

2 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192
  1. # train stage
  2. `python pdarts_train.py --data_dir '../data/' --result_path 'trial_id/result.json' --log_path 'trial_id/log' --search_space_path 'experiment_id/search_space.json' --best_selected_space_path 'experiment_id/best_selected_space.json' --trial_id 0 --model_lr 0.025 --arch_lr 3e-4 --epochs 2 --pre_epochs 1 --batch_size 64 --channels 16 --init_layers 5 --add_layer 0 6 12 --dropped_ops 3 2 1 --dropout_rates 0.1 0.4 0.7`
  3. # select stage
  4. `python pdarts_select.py --best_selected_space_path 'experiment_id/best_selected_space.json'`
  5. # retrain stage
  6. `python pdarts_retrain.py --data_dir '../data/' --result_path 'trial_id/result.json' --log_path 'trial_id/log' --best_selected_space_path 'experiment_id/best_selected_space.json' --best_checkpoint_dir 'experiment_id/' --trial_id 0 --batch_size 96 --epochs 2 --lr 0.025 --layers 20 --channels 36`
  7. # output file
  8. `result.json`
  9. ```
  10. {'type': 'Accuracy', 'result': {'sequence': 0, 'category': 'epoch', 'value': 0.1}}
  11. {'type': 'Accuracy', 'result': {'sequence': 1, 'category': 'epoch', 'value': 0.0}}
  12. {'type': 'Accuracy', 'result': {'sequence': 2, 'category': 'epoch', 'value': 0.0}}
  13. {'type': 'Accuracy', 'result': {'sequence': 3, 'category': 'epoch', 'value': 0.0}}
  14. {'type': 'Accuracy', 'result': {'sequence': 4, 'category': 'epoch', 'value': 0.0}}
  15. {'type': 'Cost_time', 'result': {'value': '41.614346981048584 s'}}
  16. ```
  17. `search_space.json`
  18. ```
  19. {
  20. "op_list": {
  21. "_type": "layer_choice",
  22. "_value": [
  23. "maxpool",
  24. "avgpool",
  25. "skipconnect",
  26. "sepconv3x3",
  27. "sepconv5x5",
  28. "dilconv3x3",
  29. "dilconv5x5",
  30. "none"
  31. ]
  32. },
  33. "search_space": {
  34. "normal_n2_p0": "op_list",
  35. "normal_n2_p1": "op_list",
  36. "normal_n2_switch": {
  37. "_type": "input_choice",
  38. "_value": {
  39. "candidates": [
  40. "normal_n2_p0",
  41. "normal_n2_p1"
  42. ],
  43. "n_chosen": 2
  44. }
  45. },
  46. ...
  47. }
  48. ```
  49. `best_selected_space.json`
  50. ```
  51. {
  52. {
  53. "op_list": {
  54. "_type": "layer_choice",
  55. "_value": [
  56. "maxpool",
  57. "avgpool",
  58. "skipconnect",
  59. "sepconv3x3",
  60. "sepconv5x5",
  61. "dilconv3x3",
  62. "dilconv5x5"
  63. ]
  64. },
  65. "best_selected_space": {
  66. "normal_n2_p0": [
  67. false,
  68. false,
  69. false,
  70. false,
  71. true,
  72. false,
  73. false
  74. ],
  75. "normal_n2_p1": [
  76. true,
  77. false,
  78. false,
  79. false,
  80. false,
  81. false,
  82. false
  83. ],
  84. ...
  85. }
  86. ```

一站式算法开发平台、高性能分布式深度学习框架、先进算法模型库、视觉模型炼知平台、数据可视化分析平台等一系列平台及工具,在模型高效分布式训练、数据处理和可视分析、模型炼知和轻量化等技术上形成独特优势,目前已在产学研等各领域近千家单位及个人提供AI应用赋能