量化自定义音频标记模型精度损失过大

1.芯片型号:X3派

2.天工开物开发包OpenExplorer版本:Ai_Toolchain_Package-release-v1.20.1-OE-v2.6.2b

3.问题定位:模型转换

4.问题具体描述:根据官方教程走的模型量化,具体模型结构:EfficientNet + Attention,音频的频谱特征提取和预处理部分放在模型外部处理,没有使用官方提供的预处理算子,校准数据也是同步的预处理过后才保存下来的特征,已检查yaml配置文件中input_type_rt、input_type_train参数均为featuremap,目前发现量化过程中部分量化节点相似度偏离严重,整体mAP下降了10几个点,推理结果也经常出现nan的情况。

==============================================================================================================================
Node                              ON   Subgraph  Type                          Cosine Similarity  Threshold  In/Out DataType  
------------------------------------------------------------------------------------------------------------------------------
Conv_2                            BPU  id(0)     HzSQuantizedConv              0.999836           3.078589   int8/int8        
Mul_4                             BPU  id(0)     HzLut                         0.999788           31.069048  int8/int8        
Conv_6                            BPU  id(0)     HzSQuantizedConv              0.994892           31.069048  int8/int8        
Mul_8                             BPU  id(0)     HzLut                         0.995812           52.429764  int8/int8        
GlobalAveragePool_9_split_id_0    BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_9_split_id_1    BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_9_split_id_2    BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_9_split_id_3    BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_9_split_id_4    BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_9_split_id_5    BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_9_split_id_6    BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_9               BPU  id(0)     HzQuantizedGlobalAveragePool  0.942759           52.429764  int8/int8        
Conv_10                           BPU  id(0)     HzSQuantizedConv              0.958868           52.429764  int8/int8        
Mul_12                            BPU  id(0)     HzLut                         0.383561           13.143455  int8/int8        
Conv_13                           BPU  id(0)     HzSQuantizedConv              0.957470           13.143430  int8/int8        
Sigmoid_14                        BPU  id(0)     HzLut                         0.992884           5.498970   int8/int8        
Mul_15                            BPU  id(0)     HzSElementwiseMul             0.983005           52.429764  int8/int8        
Conv_16                           BPU  id(0)     HzSQuantizedConv              0.967133           15.743910  int8/int8        
Conv_18                           BPU  id(0)     HzSQuantizedConv              0.928256           83.670822  int8/int8        
Mul_20                            BPU  id(0)     HzLut                         0.942507           92.699295  int8/int8        
GlobalAveragePool_21_split_id_0   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_21_split_id_1   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_21_split_id_2   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_21_split_id_3   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_21_split_id_4   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_21_split_id_5   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_21_split_id_6   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_21              BPU  id(0)     HzQuantizedGlobalAveragePool  0.838270           92.699295  int8/int8        
Conv_22                           BPU  id(0)     HzSQuantizedConv              0.960182           92.699295  int8/int8        
Mul_24                            BPU  id(0)     HzLut                         0.993296           28.492252  int8/int8        
Conv_25                           BPU  id(0)     HzSQuantizedConv              0.999187           28.492252  int8/int8        
Sigmoid_26                        BPU  id(0)     HzLut                         0.988789           5.498970   int8/int8        
Mul_27                            BPU  id(0)     HzSElementwiseMul             0.942975           92.699295  int8/int8        
Conv_28                           BPU  id(0)     HzSQuantizedConv              0.957147           15.388580  int8/int8        
Conv_30                           BPU  id(0)     HzSQuantizedConv              0.976131           80.534935  int8/int8        
Mul_32                            BPU  id(0)     HzLut                         0.977876           51.902950  int8/int8        
Conv_34                           BPU  id(0)     HzSQuantizedConv              0.901640           51.902950  int8/int8        
Mul_36                            BPU  id(0)     HzLut                         0.892430           88.654739  int8/int8        
GlobalAveragePool_37_split_id_0   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_37_split_id_1   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_37_split_id_2   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_37_split_id_3   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_37_split_id_4   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_37_split_id_5   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_37              BPU  id(0)     HzQuantizedGlobalAveragePool  0.793086           88.654739  int8/int8        
Conv_38                           BPU  id(0)     HzSQuantizedConv              0.986892           88.654739  int8/int8        
Mul_40                            BPU  id(0)     HzLut                         0.986222           15.612015  int8/int8        
Conv_41                           BPU  id(0)     HzSQuantizedConv              0.998032           15.612013  int8/int8        
Sigmoid_42                        BPU  id(0)     HzLut                         0.999481           5.498970   int8/int8        
Mul_43                            BPU  id(0)     HzSElementwiseMul             0.892153           88.654739  int8/int8        
Conv_44                           BPU  id(0)     HzSQuantizedConv              0.804178           49.186047  int8/int8        
Conv_45                           BPU  id(0)     HzSQuantizedConv              0.887929           44.279488  int8/int8        
Mul_47                            BPU  id(0)     HzLut                         0.825857           21.555767  int8/int8        
Conv_49                           BPU  id(0)     HzSQuantizedConv              0.669168           21.555767  int8/int8        
Mul_51                            BPU  id(0)     HzLut                         0.508039           38.233898  int8/int8        
GlobalAveragePool_52_split_id_0   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_52_split_id_1   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_52_split_id_2   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_52_split_id_3   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_52_split_id_4   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_52_split_id_5   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_52              BPU  id(0)     HzQuantizedGlobalAveragePool  0.221949           38.233898  int8/int8        
Conv_53                           BPU  id(0)     HzSQuantizedConv              -0.103754          38.233898  int8/int8        
Mul_55                            BPU  id(0)     HzLut                         0.213330           14.593653  int8/int8        
Conv_56                           BPU  id(0)     HzSQuantizedConv              0.399471           14.593646  int8/int8        
Sigmoid_57                        BPU  id(0)     HzLut                         0.918891           5.498970   int8/int8        
Mul_58                            BPU  id(0)     HzSElementwiseMul             0.767594           38.233898  int8/int8        
Conv_59                           BPU  id(0)     HzSQuantizedConv              0.769372           12.209419  int8/int8        
Conv_61                           BPU  id(0)     HzSQuantizedConv              0.912457           47.961449  int8/int8        
Mul_63                            BPU  id(0)     HzLut                         0.887487           22.020706  int8/int8        
Conv_65                           BPU  id(0)     HzSQuantizedConv              0.765138           22.020706  int8/int8        
Mul_67                            BPU  id(0)     HzLut                         0.662925           25.545095  int8/int8        
GlobalAveragePool_68_split_id_0   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_68_split_id_1   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_68_split_id_2   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_68_split_id_3   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_68_split_id_4   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_68_split_id_5   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_68              BPU  id(0)     HzQuantizedGlobalAveragePool  0.434824           25.545095  int8/int8        
Conv_69                           BPU  id(0)     HzSQuantizedConv              0.498449           25.545095  int8/int8        
Mul_71                            BPU  id(0)     HzLut                         0.491640           5.660278   int8/int8        
Conv_72                           BPU  id(0)     HzSQuantizedConv              0.469262           5.640640   int8/int8        
Sigmoid_73                        BPU  id(0)     HzLut                         0.925024           5.498970   int8/int8        
Mul_74                            BPU  id(0)     HzSElementwiseMul             0.616608           25.545095  int8/int8        
Conv_75                           BPU  id(0)     HzSQuantizedConv              0.804078           10.499642  int8/int8        
Conv_77                           BPU  id(0)     HzSQuantizedConv              0.909156           56.776936  int8/int8        
Mul_79                            BPU  id(0)     HzLut                         0.928730           18.397852  int8/int8        
Conv_81                           BPU  id(0)     HzSQuantizedConv              0.896441           18.397852  int8/int8        
Mul_83                            BPU  id(0)     HzLut                         0.899704           19.198795  int8/int8        
GlobalAveragePool_84_split_id_0   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_84_split_id_1   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_84_split_id_2   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_84_split_id_3   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_84_split_id_4   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_84              BPU  id(0)     HzQuantizedGlobalAveragePool  0.926736           19.198795  int8/int8        
Conv_85                           BPU  id(0)     HzSQuantizedConv              0.972449           19.198795  int8/int8        
Mul_87                            BPU  id(0)     HzLut                         0.820510           10.287921  int8/int8        
Conv_88                           BPU  id(0)     HzSQuantizedConv              0.980235           10.287570  int8/int8        
Sigmoid_89                        BPU  id(0)     HzLut                         0.999726           5.498970   int8/int8        
Mul_90                            BPU  id(0)     HzSElementwiseMul             0.903960           19.198795  int8/int8        
Conv_91                           BPU  id(0)     HzSQuantizedConv              0.807797           7.900773   int8/int8        
Conv_92                           BPU  id(0)     HzSQuantizedConv              0.905355           49.820873  int8/int8        
Mul_94                            BPU  id(0)     HzLut                         0.899956           11.497888  int8/int8        
Conv_96                           BPU  id(0)     HzSQuantizedConv              0.823604           11.497771  int8/int8        
Mul_98                            BPU  id(0)     HzLut                         0.745737           34.561031  int8/int8        
GlobalAveragePool_99_split_id_0   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_99_split_id_1   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_99_split_id_2   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_99_split_id_3   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_99_split_id_4   BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_99              BPU  id(0)     HzQuantizedGlobalAveragePool  0.547529           34.561031  int8/int8        
Conv_100                          BPU  id(0)     HzSQuantizedConv              0.358627           34.561031  int8/int8        
Mul_102                           BPU  id(0)     HzLut                         0.651810           13.437469  int8/int8        
Conv_103                          BPU  id(0)     HzSQuantizedConv              0.740769           13.437450  int8/int8        
Sigmoid_104                       BPU  id(0)     HzLut                         0.952233           5.498970   int8/int8        
Mul_105                           BPU  id(0)     HzSElementwiseMul             0.706225           34.561031  int8/int8        
Conv_106                          BPU  id(0)     HzSQuantizedConv              0.804603           10.909831  int8/int8        
Conv_108                          BPU  id(0)     HzSQuantizedConv              0.904564           58.329304  int8/int8        
Mul_110                           BPU  id(0)     HzLut                         0.899595           15.691880  int8/int8        
Conv_112                          BPU  id(0)     HzSQuantizedConv              0.834650           15.691878  int8/int8        
Mul_114                           BPU  id(0)     HzLut                         0.730096           45.668438  int8/int8        
GlobalAveragePool_115_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_115_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_115_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_115_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_115_split_id_4  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_115             BPU  id(0)     HzQuantizedGlobalAveragePool  0.016669           45.668438  int8/int8        
Conv_116                          BPU  id(0)     HzSQuantizedConv              0.615961           45.668438  int8/int8        
Mul_118                           BPU  id(0)     HzLut                         0.630204           9.503960   int8/int8        
Conv_119                          BPU  id(0)     HzSQuantizedConv              0.877433           9.503252   int8/int8        
Sigmoid_120                       BPU  id(0)     HzLut                         0.985714           5.498970   int8/int8        
Mul_121                           BPU  id(0)     HzSElementwiseMul             0.740152           45.668438  int8/int8        
Conv_122                          BPU  id(0)     HzSQuantizedConv              0.813167           8.647323   int8/int8        
Conv_124                          BPU  id(0)     HzSQuantizedConv              0.898362           61.061432  int8/int8        
Mul_126                           BPU  id(0)     HzLut                         0.884334           13.641832  int8/int8        
Conv_128                          BPU  id(0)     HzSQuantizedConv              0.903098           13.641816  int8/int8        
Mul_130                           BPU  id(0)     HzLut                         0.910590           14.174388  int8/int8        
GlobalAveragePool_131_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_131_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_131_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_131_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_131             BPU  id(0)     HzQuantizedGlobalAveragePool  0.942910           14.174377  int8/int8        
Conv_132                          BPU  id(0)     HzSQuantizedConv              0.999084           14.174377  int8/int8        
Mul_134                           BPU  id(0)     HzLut                         0.939346           8.524050   int8/int8        
Conv_135                          BPU  id(0)     HzSQuantizedConv              0.994392           8.522357   int8/int8        
Sigmoid_136                       BPU  id(0)     HzLut                         0.999963           5.498970   int8/int8        
Mul_137                           BPU  id(0)     HzSElementwiseMul             0.912133           14.174377  int8/int8        
Conv_138                          BPU  id(0)     HzSQuantizedConv              0.792652           8.429573   int8/int8        
Conv_139                          BPU  id(0)     HzSQuantizedConv              0.891801           50.787785  int8/int8        
Mul_141                           BPU  id(0)     HzLut                         0.877622           7.797539   int8/int8        
Conv_143                          BPU  id(0)     HzSQuantizedConv              0.837979           7.794337   int8/int8        
Mul_145                           BPU  id(0)     HzLut                         0.725191           31.175846  int8/int8        
GlobalAveragePool_146_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_146_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_146_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_146_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_146             BPU  id(0)     HzQuantizedGlobalAveragePool  0.637697           31.175846  int8/int8        
Conv_147                          BPU  id(0)     HzSQuantizedConv              0.840162           31.175846  int8/int8        
Mul_149                           BPU  id(0)     HzLut                         0.869121           16.964600  int8/int8        
Conv_150                          BPU  id(0)     HzSQuantizedConv              0.954371           16.964600  int8/int8        
Sigmoid_151                       BPU  id(0)     HzLut                         0.985348           5.498970   int8/int8        
Mul_152                           BPU  id(0)     HzSElementwiseMul             0.739761           31.175846  int8/int8        
Conv_153                          BPU  id(0)     HzSQuantizedConv              0.783806           7.986022   int8/int8        
Conv_155                          BPU  id(0)     HzSQuantizedConv              0.894916           48.743267  int8/int8        
Mul_157                           BPU  id(0)     HzLut                         0.870262           7.976143   int8/int8        
Conv_159                          BPU  id(0)     HzSQuantizedConv              0.838437           7.973404   int8/int8        
Mul_161                           BPU  id(0)     HzLut                         0.790434           23.697947  int8/int8        
GlobalAveragePool_162_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_162_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_162_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_162_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_162             BPU  id(0)     HzQuantizedGlobalAveragePool  0.737532           23.697947  int8/int8        
Conv_163                          BPU  id(0)     HzSQuantizedConv              0.855322           23.697947  int8/int8        
Mul_165                           BPU  id(0)     HzLut                         0.858719           14.761307  int8/int8        
Conv_166                          BPU  id(0)     HzSQuantizedConv              0.937145           14.761301  int8/int8        
Sigmoid_167                       BPU  id(0)     HzLut                         0.976754           5.498970   int8/int8        
Mul_168                           BPU  id(0)     HzSElementwiseMul             0.702142           23.697947  int8/int8        
Conv_169                          BPU  id(0)     HzSQuantizedConv              0.785340           8.671173   int8/int8        
Conv_171                          BPU  id(0)     HzSQuantizedConv              0.922907           57.211384  int8/int8        
Mul_173                           BPU  id(0)     HzLut                         0.886032           8.994512   int8/int8        
Conv_175                          BPU  id(0)     HzSQuantizedConv              0.860684           8.993396   int8/int8        
Mul_177                           BPU  id(0)     HzLut                         0.769005           25.632656  int8/int8        
GlobalAveragePool_178_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_178_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_178_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_178_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_178             BPU  id(0)     HzQuantizedGlobalAveragePool  0.653546           25.632656  int8/int8        
Conv_179                          BPU  id(0)     HzSQuantizedConv              0.877028           25.632656  int8/int8        
Mul_181                           BPU  id(0)     HzLut                         0.890879           11.833912  int8/int8        
Conv_182                          BPU  id(0)     HzSQuantizedConv              0.961302           11.833826  int8/int8        
Sigmoid_183                       BPU  id(0)     HzLut                         0.988145           5.498970   int8/int8        
Mul_184                           BPU  id(0)     HzSElementwiseMul             0.762360           25.632656  int8/int8        
Conv_185                          BPU  id(0)     HzSQuantizedConv              0.800228           8.902781   int8/int8        
Conv_187                          BPU  id(0)     HzSQuantizedConv              0.872164           55.532520  int8/int8        
Mul_189                           BPU  id(0)     HzLut                         0.868590           11.562915  int8/int8        
Conv_191                          BPU  id(0)     HzSQuantizedConv              0.824745           11.562804  int8/int8        
Mul_193                           BPU  id(0)     HzLut                         0.784823           26.881416  int8/int8        
GlobalAveragePool_194_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_194_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_194_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_194_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_194             BPU  id(0)     HzQuantizedGlobalAveragePool  0.561001           26.881416  int8/int8        
Conv_195                          BPU  id(0)     HzSQuantizedConv              0.577106           26.881416  int8/int8        
Mul_197                           BPU  id(0)     HzLut                         0.678004           11.927011  int8/int8        
Conv_198                          BPU  id(0)     HzSQuantizedConv              0.900575           11.926932  int8/int8        
Sigmoid_199                       BPU  id(0)     HzLut                         0.983859           5.498970   int8/int8        
Mul_200                           BPU  id(0)     HzSElementwiseMul             0.808966           26.881416  int8/int8        
Conv_201                          BPU  id(0)     HzSQuantizedConv              0.810349           7.465541   int8/int8        
Conv_202                          BPU  id(0)     HzSQuantizedConv              0.909702           43.018112  int8/int8        
Mul_204                           BPU  id(0)     HzLut                         0.868671           8.550163   int8/int8        
Conv_206                          BPU  id(0)     HzSQuantizedConv              0.872016           8.548510   int8/int8        
Mul_208                           BPU  id(0)     HzLut                         0.840348           22.070004  int8/int8        
GlobalAveragePool_209_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_209_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_209_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_209_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_209             BPU  id(0)     HzQuantizedGlobalAveragePool  0.697896           22.070004  int8/int8        
Conv_210                          BPU  id(0)     HzSQuantizedConv              0.906087           22.070004  int8/int8        
Mul_212                           BPU  id(0)     HzLut                         0.941245           15.490071  int8/int8        
Conv_213                          BPU  id(0)     HzSQuantizedConv              0.978055           15.490067  int8/int8        
Sigmoid_214                       BPU  id(0)     HzLut                         0.989542           5.498970   int8/int8        
Mul_215                           BPU  id(0)     HzSElementwiseMul             0.755209           22.070004  int8/int8        
Conv_216                          BPU  id(0)     HzSQuantizedConv              0.812771           6.499563   int8/int8        
Conv_218                          BPU  id(0)     HzSQuantizedConv              0.908943           43.390968  int8/int8        
Mul_220                           BPU  id(0)     HzLut                         0.847321           8.484109   int8/int8        
Conv_222                          BPU  id(0)     HzSQuantizedConv              0.855183           8.482356   int8/int8        
Mul_224                           BPU  id(0)     HzLut                         0.790803           17.407606  int8/int8        
GlobalAveragePool_225_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_225_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_225_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_225_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_225             BPU  id(0)     HzQuantizedGlobalAveragePool  0.584587           17.407606  int8/int8        
Conv_226                          BPU  id(0)     HzSQuantizedConv              0.702315           17.407606  int8/int8        
Mul_228                           BPU  id(0)     HzLut                         0.794442           11.161971  int8/int8        
Conv_229                          BPU  id(0)     HzSQuantizedConv              0.921147           11.161813  int8/int8        
Sigmoid_230                       BPU  id(0)     HzLut                         0.974040           5.498970   int8/int8        
Mul_231                           BPU  id(0)     HzSElementwiseMul             0.756021           17.407606  int8/int8        
Conv_232                          BPU  id(0)     HzSQuantizedConv              0.818496           7.391370   int8/int8        
Conv_234                          BPU  id(0)     HzSQuantizedConv              0.914589           44.851097  int8/int8        
Mul_236                           BPU  id(0)     HzLut                         0.840654           8.589934   int8/int8        
Conv_238                          BPU  id(0)     HzSQuantizedConv              0.849163           8.588337   int8/int8        
Mul_240                           BPU  id(0)     HzLut                         0.767938           18.916973  int8/int8        
GlobalAveragePool_241_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_241_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_241_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_241_split_id_3  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_241             BPU  id(0)     HzQuantizedGlobalAveragePool  0.467364           18.916973  int8/int8        
Conv_242                          BPU  id(0)     HzSQuantizedConv              0.740808           18.916973  int8/int8        
Mul_244                           BPU  id(0)     HzLut                         0.748509           13.223625  int8/int8        
Conv_245                          BPU  id(0)     HzSQuantizedConv              0.905334           13.223601  int8/int8        
Sigmoid_246                       BPU  id(0)     HzLut                         0.967348           5.498970   int8/int8        
Mul_247                           BPU  id(0)     HzSElementwiseMul             0.744530           18.916973  int8/int8        
Conv_248                          BPU  id(0)     HzSQuantizedConv              0.821513           7.374098   int8/int8        
Conv_250                          BPU  id(0)     HzSQuantizedConv              0.938354           45.916355  int8/int8        
Mul_252                           BPU  id(0)     HzLut                         0.865017           9.638214   int8/int8        
Conv_254                          BPU  id(0)     HzSQuantizedConv              0.899396           9.637586   int8/int8        
Mul_256                           BPU  id(0)     HzLut                         0.907549           15.315049  int8/int8        
GlobalAveragePool_257_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_257_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_257_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_257             BPU  id(0)     HzQuantizedGlobalAveragePool  0.967605           15.315045  int8/int8        
Conv_258                          BPU  id(0)     HzSQuantizedConv              0.999970           15.315045  int8/int8        
Mul_260                           BPU  id(0)     HzLut                         0.976281           7.535906   int8/int8        
Conv_261                          BPU  id(0)     HzSQuantizedConv              0.985517           7.531887   int8/int8        
Sigmoid_262                       BPU  id(0)     HzLut                         0.999962           5.498970   int8/int8        
Mul_263                           BPU  id(0)     HzSElementwiseMul             0.909789           15.315045  int8/int8        
Conv_264                          BPU  id(0)     HzSQuantizedConv              0.790469           5.667628   int8/int8        
Conv_265                          BPU  id(0)     HzSQuantizedConv              0.908451           37.801826  int8/int8        
Mul_267                           BPU  id(0)     HzLut                         0.854004           6.781938   int8/int8        
Conv_269                          BPU  id(0)     HzSQuantizedConv              0.874613           6.774256   int8/int8        
Mul_271                           BPU  id(0)     HzLut                         0.796914           16.740446  int8/int8        
GlobalAveragePool_272_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_272_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_272_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_272             BPU  id(0)     HzQuantizedGlobalAveragePool  0.763712           16.740446  int8/int8        
Conv_273                          BPU  id(0)     HzSQuantizedConv              0.839433           16.740446  int8/int8        
Mul_275                           BPU  id(0)     HzLut                         0.861672           16.555759  int8/int8        
Conv_276                          BPU  id(0)     HzSQuantizedConv              0.922068           16.555758  int8/int8        
Sigmoid_277                       BPU  id(0)     HzLut                         0.937180           5.498970   int8/int8        
Mul_278                           BPU  id(0)     HzSElementwiseMul             0.696421           16.740446  int8/int8        
Conv_279                          BPU  id(0)     HzSQuantizedConv              0.789078           4.012821   int8/int8        
Conv_281                          BPU  id(0)     HzSQuantizedConv              0.914476           40.901840  int8/int8        
Mul_283                           BPU  id(0)     HzLut                         0.852786           6.736721   int8/int8        
Conv_285                          BPU  id(0)     HzSQuantizedConv              0.858286           6.728738   int8/int8        
Mul_287                           BPU  id(0)     HzLut                         0.750518           16.346664  int8/int8        
GlobalAveragePool_288_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_288_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_288_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_288             BPU  id(0)     HzQuantizedGlobalAveragePool  0.706490           16.346663  int8/int8        
Conv_289                          BPU  id(0)     HzSQuantizedConv              0.748821           16.346663  int8/int8        
Mul_291                           BPU  id(0)     HzLut                         0.819174           15.751864  int8/int8        
Conv_292                          BPU  id(0)     HzSQuantizedConv              0.892673           15.751863  int8/int8        
Sigmoid_293                       BPU  id(0)     HzLut                         0.975658           5.498970   int8/int8        
Mul_294                           BPU  id(0)     HzSElementwiseMul             0.714289           16.346663  int8/int8        
Conv_295                          BPU  id(0)     HzSQuantizedConv              0.793221           6.081708   int8/int8        
Conv_297                          BPU  id(0)     HzSQuantizedConv              0.917334           44.239944  int8/int8        
Mul_299                           BPU  id(0)     HzLut                         0.833319           6.887061   int8/int8        
Conv_301                          BPU  id(0)     HzSQuantizedConv              0.842434           6.880036   int8/int8        
Mul_303                           BPU  id(0)     HzLut                         0.743629           15.282858  int8/int8        
GlobalAveragePool_304_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_304_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_304_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_304             BPU  id(0)     HzQuantizedGlobalAveragePool  0.729354           15.282854  int8/int8        
Conv_305                          BPU  id(0)     HzSQuantizedConv              0.809021           15.282854  int8/int8        
Mul_307                           BPU  id(0)     HzLut                         0.838657           11.609028  int8/int8        
Conv_308                          BPU  id(0)     HzSQuantizedConv              0.898425           11.608923  int8/int8        
Sigmoid_309                       BPU  id(0)     HzLut                         0.982799           5.498970   int8/int8        
Mul_310                           BPU  id(0)     HzSElementwiseMul             0.694917           15.282854  int8/int8        
Conv_311                          BPU  id(0)     HzSQuantizedConv              0.793641           5.843131   int8/int8        
Conv_313                          BPU  id(0)     HzSQuantizedConv              0.925385           44.628574  int8/int8        
Mul_315                           BPU  id(0)     HzLut                         0.826158           7.986450   int8/int8        
Conv_317                          BPU  id(0)     HzSQuantizedConv              0.838098           7.983736   int8/int8        
Mul_319                           BPU  id(0)     HzLut                         0.730630           16.378695  int8/int8        
GlobalAveragePool_320_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_320_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_320_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_320             BPU  id(0)     HzQuantizedGlobalAveragePool  0.691544           16.378693  int8/int8        
Conv_321                          BPU  id(0)     HzSQuantizedConv              0.702137           16.378693  int8/int8        
Mul_323                           BPU  id(0)     HzLut                         0.728958           8.522956   int8/int8        
Conv_324                          BPU  id(0)     HzSQuantizedConv              0.830651           8.521261   int8/int8        
Sigmoid_325                       BPU  id(0)     HzLut                         0.980579           5.498970   int8/int8        
Mul_326                           BPU  id(0)     HzSElementwiseMul             0.732954           16.378693  int8/int8        
Conv_327                          BPU  id(0)     HzSQuantizedConv              0.804464           7.871977   int8/int8        
Conv_329                          BPU  id(0)     HzSQuantizedConv              0.946942           52.423195  int8/int8        
Mul_331                           BPU  id(0)     HzLut                         0.914441           7.704774   int8/int8        
Conv_333                          BPU  id(0)     HzSQuantizedConv              0.881819           7.701303   int8/int8        
Mul_335                           BPU  id(0)     HzLut                         0.749431           19.437279  int8/int8        
GlobalAveragePool_336_split_id_0  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_336_split_id_1  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_336_split_id_2  BPU  id(0)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_336             BPU  id(0)     HzQuantizedGlobalAveragePool  0.668114           19.437279  int8/int8        
Conv_337                          BPU  id(0)     HzSQuantizedConv              0.549668           19.437279  int8/int8        
Mul_339                           BPU  id(0)     HzLut                         0.638028           12.301728  int8/int8        
Conv_340                          BPU  id(0)     HzSQuantizedConv              0.724144           12.301673  int8/int8        
Sigmoid_341                       BPU  id(0)     HzLut                         0.900135           5.498970   int8/int8        
Mul_342                           BPU  id(0)     HzSElementwiseMul             0.701035           19.437279  int8/int8        
Conv_343                          BPU  id(0)     HzSQuantizedConv              0.758191           14.395515  int8/int8        
Conv_344_sub1                     BPU  id(0)     HzSQuantizedConv              0.919195           23.280869  int8/int8        
Conv_344_sub2                     BPU  id(0)     HzSQuantizedConv              0.921153           23.280869  int8/int8        
Conv_344_concat                   BPU  id(0)     Concat                        0.920285           5.751357   int8/int8        
Mul_346                           BPU  id(0)     HzLut                         0.844131           5.751357   int8/int8        
Conv_348                          CPU  --        Conv                          0.857950           5.733134   float/float      
Mul_350                           BPU  id(1)     HzLut                         0.757617           12.446754  int8/int8        
GlobalAveragePool_351_split_id_0  BPU  id(1)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_351_split_id_1  BPU  id(1)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_351_split_id_2  BPU  id(1)     HzQuantizedAveragePool                                      int8/int8        
GlobalAveragePool_351             BPU  id(1)     HzQuantizedGlobalAveragePool  0.768771           12.446706  int8/int8        
Conv_352                          BPU  id(1)     HzSQuantizedConv              0.959496           12.446706  int8/int8        
Mul_354                           BPU  id(1)     HzLut                         0.951071           7.105406   int8/int8        
Conv_355                          BPU  id(1)     HzSQuantizedConv              0.980208           7.099580   int8/int8        
Sigmoid_356                       BPU  id(1)     HzLut                         0.998196           5.498970   int8/int8        
Mul_357                           BPU  id(1)     HzSElementwiseMul             0.721501           12.446706  int8/int8        
Conv_358_split                    BPU  id(1)     Split                                                       int8/int8        
Conv_358_sub1                     BPU  id(1)     HzSQuantizedConv              0.822564           7.949086   int8/int8        
Conv_358_sub2                     BPU  id(1)     HzSQuantizedConv              0.810408           7.949086   int8/int8        
UNIT_CONV_FOR_Add_359             BPU  id(1)     HzSQuantizedConv              0.765254           18.327040  int8/int8        
Conv_360                          BPU  id(1)     HzSQuantizedConv              0.866615           32.128651  int8/int8        
Mul_362                           BPU  id(1)     HzLut                         0.762793           22.066324  int8/int8        
AveragePool_364                   BPU  id(1)     HzSQuantizedConv              0.714611           22.066324  int8/int32       
Conv_366                          BPU  id(2)     HzSQuantizedConv              0.881503           7.171721   int8/int8        
Sigmoid_367                       BPU  id(2)     HzLut                         0.834804           5.498970   int8/int8        
Conv_368                          BPU  id(2)     HzSQuantizedConv              0.920116           7.171721   int8/int8        
Sigmoid_369                       BPU  id(2)     HzLut                         0.687944           5.498970   int8/int8        
Gather_371                        CPU  --        Gather                        0.834805           0.995926   float/float      
Gather_373                        CPU  --        Gather                        0.687951           0.995926   float/float      
Gather_373_reshape                CPU  --        Reshape                                                     float/float      
Clip_374                          BPU  id(3)     HzLut                         0.834804           0.995926   int8/int8        
Clip_374_reshape                  CPU  --        Reshape                                                     float/float      
ReduceSum_375                     CPU  --        ReduceSum                     0.935969           --         float/float      
Div_377_reciprocal                CPU  --        Reciprocal                    0.555402           --         float/float      
Div_377_mul                       CPU  --        Mul                           0.768257           --         float/float      
Mul_378                           CPU  --        Mul                           0.693240           --         float/float      
ReduceSum_379                     CPU  --        ReduceSum                     0.839462           --         float/float      
ReduceSum_379_reshape             CPU  --        Reshape                                                     float/float      
Mul_380                           CPU  --        Mul                           0.839484           --         float/float      
Conv_381                          BPU  id(2)     HzSQuantizedConv              0.878739           7.171721   int8/int8        
Sigmoid_382                       BPU  id(2)     HzLut                         0.808690           5.498970   int8/int8        
Conv_383                          BPU  id(2)     HzSQuantizedConv              0.916833           7.171721   int8/int8        
Sigmoid_384                       BPU  id(2)     HzLut                         0.695682           5.498970   int8/int8        
Gather_386                        CPU  --        Gather                        0.808691           0.995926   float/float      
Gather_388                        CPU  --        Gather                        0.695690           0.995926   float/float      
Gather_388_reshape                CPU  --        Reshape                                                     float/float      
Clip_389                          BPU  id(4)     HzLut                         0.808690           0.995926   int8/int8        
Clip_389_reshape                  CPU  --        Reshape                                                     float/float      
ReduceSum_390                     CPU  --        ReduceSum                     0.908368           --         float/float      
Div_392_reciprocal                CPU  --        Reciprocal                    0.933148           --         float/float      
Div_392_mul                       CPU  --        Mul                           0.790100           --         float/float      
Mul_393                           CPU  --        Mul                           0.856828           --         float/float      
ReduceSum_394                     CPU  --        ReduceSum                     0.924828           --         float/float      
ReduceSum_394_reshape             CPU  --        Reshape                                                     float/float      
Mul_395                           CPU  --        Mul                           0.924831           --         float/float      
Conv_396                          BPU  id(2)     HzSQuantizedConv              0.894938           7.171721   int8/int8        
Sigmoid_397                       BPU  id(2)     HzLut                         0.830071           5.498970   int8/int8        
Conv_398                          BPU  id(2)     HzSQuantizedConv              0.921553           7.171721   int8/int8        
Sigmoid_399                       BPU  id(2)     HzLut                         0.657382           5.498970   int8/int8        
Gather_401                        CPU  --        Gather                        0.830072           0.995926   float/float      
Gather_403                        CPU  --        Gather                        0.657391           0.995926   float/float      
Gather_403_reshape                CPU  --        Reshape                                                     float/float      
Clip_404                          BPU  id(5)     HzLut                         0.830071           0.995926   int8/int8        
Clip_404_reshape                  CPU  --        Reshape                                                     float/float      
ReduceSum_405                     CPU  --        ReduceSum                     0.940781           --         float/float      
Div_407_reciprocal                CPU  --        Reciprocal                    0.830283           --         float/float      
Div_407_mul                       CPU  --        Mul                           0.848577           --         float/float      
Mul_408                           CPU  --        Mul                           0.769717           --         float/float      
ReduceSum_409                     CPU  --        ReduceSum                     0.856496           --         float/float      
ReduceSum_409_reshape             CPU  --        Reshape                                                     float/float      
Mul_410                           CPU  --        Mul                           0.856509           --         float/float      
Conv_411                          BPU  id(2)     HzSQuantizedConv              0.869522           7.171721   int8/int8        
Sigmoid_412                       BPU  id(2)     HzLut                         0.831501           5.498970   int8/int8        
Conv_413                          BPU  id(2)     HzSQuantizedConv              0.914767           7.171721   int8/int8        
Sigmoid_414                       BPU  id(2)     HzLut                         0.700347           5.498970   int8/int8        
Gather_416                        CPU  --        Gather                        0.831501           0.995926   float/float      
Gather_418                        CPU  --        Gather                        0.700354           0.995926   float/float      
Gather_418_reshape                CPU  --        Reshape                                                     float/float      
Clip_419                          BPU  id(6)     HzLut                         0.831500           0.995926   int8/int8        
Clip_419_reshape                  CPU  --        Reshape                                                     float/float      
ReduceSum_420                     CPU  --        ReduceSum                     0.934335           --         float/float      
Div_422_reciprocal                CPU  --        Reciprocal                    0.490855           --         float/float      
Div_422_mul                       CPU  --        Mul                           0.777003           --         float/float      
Mul_423                           CPU  --        Mul                           0.744603           --         float/float      
ReduceSum_424                     CPU  --        ReduceSum                     0.922677           --         float/float      
ReduceSum_424_reshape             CPU  --        Reshape                                                     float/float      
Mul_425                           CPU  --        Mul                           0.922682           --         float/float      
Unsqueeze_426                     CPU  --        Reshape                                                     float/float      
Unsqueeze_427                     CPU  --        Reshape                                                     float/float      
Unsqueeze_428                     CPU  --        Reshape                                                     float/float      
Unsqueeze_429                     CPU  --        Reshape                                                     float/float      
Concat_430                        CPU  --        Concat                        0.898418           --         float/float      
ReduceSum_431                     CPU  --        ReduceSum                     0.909389           --         float/float      
ReduceSum_431_reshape             CPU  --        Reshape                                                     float/float      
Sigmoid_432                       CPU  --        Sigmoid                       0.999456           --         float/float

您好,地平线工具链在持续迭代优化,为了给您提供更好的服务,希望您能抽出3分钟左右的时间,将您在使用工具链期间的感受和建议告诉我们,您的宝贵意见对我们很重要,非常感谢!-
问卷链接:地平线算法工具链使用满意度反馈

您好,您可以通过以下方式来进行PTQ的精度调优:1.尝试不同的calibration_type,使用max校准并尝试不同的分位值max_percentile;2.开启per_channel量化;3.对余弦相似度低的算子强制运行在cpu上(run_on_cpu)

你好,强制算子在cpu运行是直接在config.yaml的run_on_cpu里指定吗,还需要其他操作吗,我添加了之后会导致上个bpu算子的输出出错

请问这个是什么问题呢,强制cpu运行就会出现

方便用百度网盘提供下转换前的onnx文件和yaml吗

能私发吗,涉及到公司业务

已收到,我们分析一下

本身就运行在cpu的算子,不需要run_on_cpu,另外配置的格式手册里没有写的特别明白,是这样的-
run_on_cpu:‘Conv_69;Conv_72’-
另外我这里出现的转换日志和你那张截图不太一样,确实一下是不是在2.6.2对应的docker环境运行的呢?

2.6.2的docker环境在哪呢,我这边在docker hub上没看到

地平线 XJ3 芯片工具链 版本发布及Filezilla使用教程 (horizon.ai) 在这里哈,可以收藏下这个帖子

好的谢谢

你配的是哪些算子?把run_on_cpu后的内容贴一下吧

Mul_12;GlobalAveragePool_52;Mul_67;GlobalAveragePool_68;Conv_69;Mul_71;Conv_72;Mul_74;Mul_98;GlobalAveragePool_99;Conv_100;Mul_105;Mul_114;GlobalAveragePool_115;Conv_116;Mul_118;Mul_121;GlobalAveragePool_146;Conv_147;Mul_149;Mul_152;GlobalAveragePool_162;Conv_163;Mul_165;Mul_168;Mul_177;GlobalAveragePool_178;Mul_184;GlobalAveragePool_194;Conv_195;Mul_197;GlobalAveragePool_209;GlobalAveragePool_225;Conv_226;Mul_228;Mul_231;GlobalAveragePool_241;Conv_242;Mul_247;GlobalAveragePool_272;Mul_278;GlobalAveragePool_288;Conv_289;Mul_291;Mul_294;GlobalAveragePool_304;Conv_305;Mul_307;Mul_310;Mul_319;GlobalAveragePool_320;Conv_321;Mul_323;Conv_324;Mul_326;Mul_335;GlobalAveragePool_336;Conv_337;Mul_339;Mul_342;Mul_350;GlobalAveragePool_351;Mul_357;Conv_358_sub1;Conv_358_sub2;Mul_362;AveragePool_364;Sigmoid_367;Clip_374;Sigmoid_382;Sigmoid_384;Clip_389;Sigmoid_397;Sigmoid_399;Clip_404;Conv_411;Sigmoid_412;Sigmoid_414;Clip_419

我这里报的错和你的不一样:

ERROR *** ERROR-OCCUR-DURING {runtime.runtime_model_generation} ***, error message: HorizonRT not support these cpu operators: HzSwish

不需要配置这么多算子都run_on_bpu哈,建议试一下我们的PTQ精度debug工具,尝试一下相关调优措施,参考链接:

4.1.2.11. 精度debug工具 — Horizon Open Explorer

【PTQ精度debug示例】mnasnet_1.0_96精度问题分析 (horizon.cc)

【PTQ精度debug示例】MobileVit_s精度问题分析 (horizon.cc)

【PTQ精度debug示例】repvgg_b2_deploy精度问题分析 (horizon.cc)

另外还有配置int16量化的方法:

PTQ精度调优手段—设置Int16量化 (horizon.cc)

我想问下校准数据一定要覆盖所有类别所有场景而且数量均衡吗

校准数据别用很特殊的图(纯黑纯白,没有目标区域等)就行,没有特别严格的要求

您好,我使用了精度debug工具,普通节点、权重节点和激活节点的量化误差都不明显,但量化过程中很多算子误差很大,强制cpu运行的方法始终有报错,能帮我分析一下这些算子run_on_cpu的报错吗:

GlobalAveragePool_52;GlobalAveragePool_68;Conv_69;Conv_72;Mul_74;GlobalAveragePool_99;Conv_100;Mul_105;GlobalAveragePool_115;Conv_116;Mul_121;GlobalAveragePool_146;Conv_147;Mul_152;GlobalAveragePool_162;Conv_163;Mul_168;GlobalAveragePool_178;Mul_184;GlobalAveragePool_194;Conv_195;GlobalAveragePool_209;GlobalAveragePool_225;Conv_226;Mul_231;GlobalAveragePool_241;Conv_242;Mul_247;GlobalAveragePool_272;Mul_278;GlobalAveragePool_288;Conv_289;Mul_294;GlobalAveragePool_304;Conv_305;Mul_310;GlobalAveragePool_320;Conv_321;Conv_324;Mul_326;GlobalAveragePool_336;Conv_337;Mul_342;GlobalAveragePool_351;Mul_357;Conv_358_sub1;Conv_358_sub2;AveragePool_364;Sigmoid_367;Clip_374;Sigmoid_382;Sigmoid_384;Clip_389;Sigmoid_397;Sigmoid_399;Clip_404;Conv_411;Sigmoid_412;Sigmoid_414;Clip_419

。。。建议你可以先配置一部分算子run_on_cpu,看有没有报错,没有的话就再加一些算子,这么重复几次,出现报错了就知道是哪一个或者哪几个算子的问题了~