RDK X5双目估计算法量化后输出为纯色问题

你好,我最近在研究地平线双目深度估计相关算法,我在StereoNet的基础上进行了修改与训练,在PC端测试精度不错,然后我按照要求导出为符合X5部署的onnx格式,之后我使用hb_mapper makertbin进行量化,我的配置文件如下:

# 模型转化相关的参数
model_parameters:
  onnx_model: 'onnx/dif_data.onnx'
  march: "bayes-e"
  layer_out_dump: False
  working_dir: 'model_output_dif_data'
  output_model_file_prefix: 'model_output_dif_data'
  # remove_node_type: "Dequantize"

  node_info: {
      #  "/spx_conv3x3/spx_conv3x3.0/Conv":{
      #   "ON": "BPU",
      #   "InputType": "int16",
      #   "OutputType": "int16"
      # },                                       
      # "/spx/conv/ConvTranspose":{
      #   "ON": "BPU",
      #   "InputType": "int16",
      #   "OutputType": "int16"
      # },
      "/softmax/Softmax": {
        "ON": "BPU",
        "InputType": "int16",
        "OutputType": "int16"
      }, 
      "/softmax2/Softmax": {
        "ON": "BPU",
       "InputType": "int16",
       "OutputType": "int16"
      },                                                              
  }

input_parameters:
  # 双目 = 两个输入,用分号分隔                                               
  input_name: "left;right"                 
  input_shape: '1x3x384x640;1x3x384x640'   # 根据实际尺寸改                 
                                                                              
  # 运行时输入格式 - NV12                                                   
  input_type_rt: 'nv12;nv12'                                                  
  input_layout_rt: ''                                                         
                
  # 训练时格式 - PyTorch一般是RGB + NCHW                                      
  input_type_train: 'yuv444;yuv444'
  input_layout_train: 'NCHW;NCHW'                                             
  
  # 预处理 - 根据训练时normalization设置                                      
  norm_type: 'data_mean_and_scale;data_mean_and_scale'
                                                                              
  # ImageNet常用值,换成你训练时实际用的                                      
  mean_value: '128.0 128.0 128.0;128.0 128.0 128.0'
  scale_value: '0.0078125 0.0078125 0.0078125;0.0078125 0.0078125 0.0078125'         




# 模型量化相关参数
calibration_parameters:                                                                                                           
  cal_data_dir: './calibration_data_640x384_split/left_float32;./calibration_data_640x384_split/right_float32'                                                
  cal_data_type: 'float32;float32'                                                                                                
  calibration_type: 'default'  


# 编译器相关参数
compiler_parameters:
  compile_mode: 'latency'
  debug: False
  optimize_level: 'O3'


在我量化之后,精度如下,感觉还可以接受:

2026-05-27 10:59:03,244 INFO consumed time 104.152
2026-05-27 10:59:05,162 INFO FPS=81.56, latency = 12260.3 us, DDR = 50798976 bytes   (see main_graph_subgraph_0.html)
2026-05-27 10:59:05,287 INFO End to compile the model with march bayes-e.
2026-05-27 10:59:09,596 INFO The main quantized node information:
======================================================================================================================================
Node                                                ON   Subgraph  Type                       Cosine Similarity  Threshold  DataType  
--------------------------------------------------------------------------------------------------------------------------------------
HZ_PREPROCESS_FOR_left                              BPU  id(0)     HzSQuantizedPreprocess     1.000010           127.0      int8      
/backbone/mod1/mod1.0/Conv                          BPU  id(0)     HzSQuantizedConv           0.999939           1.00781    int8      
...one/mod2.0/head_layer/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           1.000513           3.16234    int8      
...one/mod2.0/head_layer/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999743           20.213     int8      
.../mod3.0/head_layer/downsample/downsample.0/Conv  BPU  id(0)     HzSQuantizedConv           1.000003           12.9148    int8      
....0/head_layer/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...one/mod3.0/head_layer/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999715           12.9148    int8      
...one/mod3.0/head_layer/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999662           14.7373    int8      
variable_2032_pad                                   BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999825           1.8658     int8      
...layers/stack_layers.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999357           20.313     int8      
variable_2034_pad                                   BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.1/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999517           6.15321    int8      
...layers/stack_layers.1/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999551           14.8254    int8      
.../mod4.0/head_layer/downsample/downsample.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999486           21.1246    int8      
...tack_layers.1/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...one/mod4.0/head_layer/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999556           21.1246    int8      
...one/mod4.0/head_layer/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999498           4.70103    int8      
....0/head_layer/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999643           6.89772    int8      
...layers/stack_layers.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999254           3.69102    int8      
...tack_layers.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.1/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999176           5.73961    int8      
...layers/stack_layers.1/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998649           2.67097    int8      
.../mod5.0/head_layer/downsample/downsample.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999706           4.93636    int8      
...tack_layers.1/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...one/mod5.0/head_layer/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998299           4.93636    int8      
...one/mod5.0/head_layer/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998669           2.3388     int8      
....0/head_layer/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998453           3.45497    int8      
...layers/stack_layers.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.997999           1.64744    int8      
...tack_layers.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.1/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996724           3.21956    int8      
...layers/stack_layers.1/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.992305           1.01426    int8      
...tack_layers.1/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.2/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.992960           3.17622    int8      
...layers/stack_layers.2/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988797           0.935316   int8      
...tack_layers.2/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.3/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988502           3.45553    int8      
...layers/stack_layers.3/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.980835           1.15998    int8      
...tack_layers.3/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.4/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.989714           3.2799     int8      
...layers/stack_layers.4/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988494           2.32017    int8      
...tack_layers.4/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.5/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.986282           3.80577    int8      
...layers/stack_layers.5/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.987795           2.00013    int8      
.../mod6.0/head_layer/downsample/downsample.0/Conv  BPU  id(0)     HzSQuantizedConv           0.987104           4.53167    int8      
...tack_layers.5/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...one/mod6.0/head_layer/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.991675           4.53167    int8      
...one/mod6.0/head_layer/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.992957           1.47086    int8      
....0/head_layer/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.992260           1.29606    int8      
...layers/stack_layers.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.990830           1.08635    int8      
...tack_layers.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...layers/stack_layers.1/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988902           1.24994    int8      
...layers/stack_layers.1/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995751           2.56139    int8      
HZ_PREPROCESS_FOR_right                             BPU  id(0)     HzSQuantizedPreprocess     1.000008           127.0      int8      
/backbone/mod1/mod1.0_1/Conv                        BPU  id(0)     HzSQuantizedConv           0.999871           1.00781    int8      
...e/mod2.0/head_layer/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           1.000089           3.16233    int8      
...e/mod2.0/head_layer/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999709           20.2125    int8      
...od3.0/head_layer/downsample/downsample.0_1/Conv  BPU  id(0)     HzSQuantizedConv           1.000033           12.9147    int8      
.../head_layer/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...e/mod3.0/head_layer/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999707           12.9147    int8      
...e/mod3.0/head_layer/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999632           14.7353    int8      
variable_2027_pad                                   BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.0/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999942           1.87575    int8      
...yers/stack_layers.0/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999350           20.3127    int8      
variable_2029_pad                                   BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.1/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999474           6.15515    int8      
...yers/stack_layers.1/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999555           14.8144    int8      
...od4.0/head_layer/downsample/downsample.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999174           21.1366    int8      
...ck_layers.1/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...e/mod4.0/head_layer/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999430           21.1366    int8      
...e/mod4.0/head_layer/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999437           4.70194    int8      
.../head_layer/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.0/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999454           6.89704    int8      
...yers/stack_layers.0/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.998681           3.69054    int8      
...ck_layers.0/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.1/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.998550           5.74449    int8      
...yers/stack_layers.1/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.997659           2.66851    int8      
...od5.0/head_layer/downsample/downsample.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.999568           4.9439     int8      
...ck_layers.1/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...e/mod5.0/head_layer/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.997295           4.9439     int8      
...e/mod5.0/head_layer/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.997791           2.34272    int8      
.../head_layer/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.0/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.997596           3.45775    int8      
...yers/stack_layers.0/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.996349           1.64853    int8      
...ck_layers.0/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.1/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.994415           3.22268    int8      
...yers/stack_layers.1/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.985096           1.01496    int8      
...ck_layers.1/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.2/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.986177           3.17917    int8      
...yers/stack_layers.2/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.978390           0.935306   int8      
...ck_layers.2/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.3/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.978040           3.45267    int8      
...yers/stack_layers.3/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.970643           1.1613     int8      
...ck_layers.3/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.4/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.983455           3.27744    int8      
...yers/stack_layers.4/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.982294           2.32018    int8      
...ck_layers.4/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.5/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.978830           3.80574    int8      
...yers/stack_layers.5/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.982460           2.00049    int8      
...od6.0/head_layer/downsample/downsample.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.978012           4.53231    int8      
...ck_layers.5/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...e/mod6.0/head_layer/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.982281           4.53231    int8      
...e/mod6.0/head_layer/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.983850           1.4717     int8      
.../head_layer/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.0/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.983704           1.29648    int8      
...yers/stack_layers.0/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.981720           1.08843    int8      
...ck_layers.0/relu_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...yers/stack_layers.1/conv/conv.0/conv.0.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.977779           1.24973    int8      
...yers/stack_layers.1/conv/conv.1/conv.1.0_1/Conv  BPU  id(0)     HzSQuantizedConv           0.992566           2.562      int8      
/neck/conv_extract.0/conv_extract.0.0/Conv          BPU  id(0)     HzSQuantizedConv           0.998453           4.83495    int8      
/neck/conv1x1_up.0/conv1x1_up.0.0/Conv              BPU  id(0)     HzSQuantizedConv           0.998386           6.84539    int8      
/neck/upsample.0/ConvTranspose                      BPU  id(0)     HzSQuantizedConvTranspose  0.998895           6.14636    int8      
/neck/conv_extract.1/conv_extract.1.0/Conv          BPU  id(0)     HzSQuantizedConv           0.997089           4.53167    int8      
/neck/conv1x1_up.1/conv1x1_up.1.0/Conv              BPU  id(0)     HzSQuantizedConv           0.987788           9.6324     int8      
/neck/upsample.1/ConvTranspose                      BPU  id(0)     HzSQuantizedConvTranspose  0.982623           2.30813    int8      
/neck/conv_extract.2/conv_extract.2.0/Conv          BPU  id(0)     HzSQuantizedConv           0.998159           4.93636    int8      
/neck/conv_extract.0/conv_extract.0.0_1/Conv        BPU  id(0)     HzSQuantizedConv           0.997301           4.83486    int8      
/neck/conv1x1_up.0/conv1x1_up.0.0_1/Conv            BPU  id(0)     HzSQuantizedConv           0.996895           6.84394    int8      
/neck/upsample.0_1/ConvTranspose                    BPU  id(0)     HzSQuantizedConvTranspose  0.997922           6.14634    int8      
/neck/conv_extract.1/conv_extract.1.0_1/Conv        BPU  id(0)     HzSQuantizedConv           0.994330           4.53231    int8      
/neck/conv1x1_up.1/conv1x1_up.1.0_1/Conv            BPU  id(0)     HzSQuantizedConv           0.976296           9.63273    int8      
/neck/upsample.1_1/ConvTranspose                    BPU  id(0)     HzSQuantizedConvTranspose  0.965549           2.30742    int8      
/neck/conv_extract.2/conv_extract.2.0_1/Conv        BPU  id(0)     HzSQuantizedConv           0.996129           4.9439     int8      
/Mul                                                BPU  id(0)     HzSElementwiseMul          0.999142           6.84539    int8      
/ReduceMean                                         BPU  id(0)     HzSQuantizedReduceMean     0.999095           46.8494    int8      
/Slice                                              BPU  id(0)     Slice                      0.997976           6.84539    int8      
/Slice_1                                            BPU  id(0)     Slice                      0.997401           6.84394    int8      
/Mul_1                                              BPU  id(0)     HzSElementwiseMul          0.998347           6.84539    int8      
/ReduceMean_1                                       BPU  id(0)     HzSQuantizedReduceMean     0.999327           19.5509    int8      
/gc_pad.1/Pad                                       BPU  id(0)     HzPad                      0.999328           3.29574    int8      
/Slice_2                                            BPU  id(0)     Slice                      0.998169           6.84539    int8      
/Slice_3                                            BPU  id(0)     Slice                      0.997752           6.84394    int8      
/Mul_2                                              BPU  id(0)     HzSElementwiseMul          0.999304           6.84539    int8      
/ReduceMean_2                                       BPU  id(0)     HzSQuantizedReduceMean     0.999131           16.9378    int8      
/gc_pad.2/Pad                                       BPU  id(0)     HzPad                      0.999131           3.29574    int8      
/Slice_4                                            BPU  id(0)     Slice                      0.998151           6.84539    int8      
/Slice_5                                            BPU  id(0)     Slice                      0.998045           6.84394    int8      
/Mul_3                                              BPU  id(0)     HzSElementwiseMul          0.999309           6.84539    int8      
/ReduceMean_3                                       BPU  id(0)     HzSQuantizedReduceMean     0.999372           16.5533    int8      
/gc_pad.3/Pad                                       BPU  id(0)     HzPad                      0.999372           3.29574    int8      
/Slice_6                                            BPU  id(0)     Slice                      0.998130           6.84539    int8      
/Slice_7                                            BPU  id(0)     Slice                      0.998093           6.84394    int8      
/Mul_4                                              BPU  id(0)     HzSElementwiseMul          0.999187           6.84539    int8      
/ReduceMean_4                                       BPU  id(0)     HzSQuantizedReduceMean     0.999232           16.1996    int8      
/gc_pad.4/Pad                                       BPU  id(0)     HzPad                      0.999232           3.29574    int8      
/Slice_8                                            BPU  id(0)     Slice                      0.998101           6.84539    int8      
/Slice_9                                            BPU  id(0)     Slice                      0.998095           6.84394    int8      
/Mul_5                                              BPU  id(0)     HzSElementwiseMul          0.999257           6.84539    int8      
/ReduceMean_5                                       BPU  id(0)     HzSQuantizedReduceMean     0.999229           16.211     int8      
/gc_pad.5/Pad                                       BPU  id(0)     HzPad                      0.999229           3.29574    int8      
/Concat                                             BPU  id(0)     Concat                     0.998535           3.29574    int8      
/Mul_6                                              BPU  id(0)     HzSElementwiseMul          0.998240           9.6324     int8      
/ReduceMean_6                                       BPU  id(0)     HzSQuantizedReduceMean     0.998126           92.7829    int8      
/Slice_10                                           BPU  id(0)     Slice                      0.997595           9.6324     int8      
/Slice_11                                           BPU  id(0)     Slice                      0.995898           9.63273    int8      
/Mul_7                                              BPU  id(0)     HzSElementwiseMul          0.998561           9.6324     int8      
/ReduceMean_7                                       BPU  id(0)     HzSQuantizedReduceMean     0.998283           54.6915    int8      
/gc_pad.7/Pad                                       BPU  id(0)     HzPad                      0.998283           4.77222    int8      
/Slice_12                                           BPU  id(0)     Slice                      0.997492           9.6324     int8      
/Slice_13                                           BPU  id(0)     Slice                      0.995955           9.63273    int8      
/Mul_8                                              BPU  id(0)     HzSElementwiseMul          0.998440           9.6324     int8      
/ReduceMean_8                                       BPU  id(0)     HzSQuantizedReduceMean     0.998939           42.3482    int8      
/gc_pad.8/Pad                                       BPU  id(0)     HzPad                      0.998939           4.77222    int8      
/Slice_14                                           BPU  id(0)     Slice                      0.997458           9.6324     int8      
/Slice_15                                           BPU  id(0)     Slice                      0.996047           9.63273    int8      
/Mul_9                                              BPU  id(0)     HzSElementwiseMul          0.998606           9.6324     int8      
/ReduceMean_9                                       BPU  id(0)     HzSQuantizedReduceMean     0.998192           38.8949    int8      
/gc_pad.9/Pad                                       BPU  id(0)     HzPad                      0.998192           4.77222    int8      
/Slice_16                                           BPU  id(0)     Slice                      0.997476           9.6324     int8      
/Slice_17                                           BPU  id(0)     Slice                      0.996125           9.63273    int8      
/Mul_10                                             BPU  id(0)     HzSElementwiseMul          0.998659           9.6324     int8      
/ReduceMean_10                                      BPU  id(0)     HzSQuantizedReduceMean     0.999192           38.2187    int8      
/gc_pad.10/Pad                                      BPU  id(0)     HzPad                      0.999192           4.77222    int8      
/Slice_18                                           BPU  id(0)     Slice                      0.997479           9.6324     int8      
/Slice_19                                           BPU  id(0)     Slice                      0.996187           9.63273    int8      
/Mul_11                                             BPU  id(0)     HzSElementwiseMul          0.998687           9.6324     int8      
/ReduceMean_11                                      BPU  id(0)     HzSQuantizedReduceMean     0.998446           38.5086    int8      
/gc_pad.11/Pad                                      BPU  id(0)     HzPad                      0.998446           4.77222    int8      
/Slice_20                                           BPU  id(0)     Slice                      0.997473           9.6324     int8      
/Slice_21                                           BPU  id(0)     Slice                      0.996208           9.63273    int8      
/Mul_12                                             BPU  id(0)     HzSElementwiseMul          0.998599           9.6324     int8      
/ReduceMean_12                                      BPU  id(0)     HzSQuantizedReduceMean     0.999204           38.4273    int8      
/gc_pad.12/Pad                                      BPU  id(0)     HzPad                      0.999204           4.77222    int8      
/Slice_22                                           BPU  id(0)     Slice                      0.997469           9.6324     int8      
/Slice_23                                           BPU  id(0)     Slice                      0.996236           9.63273    int8      
/Mul_13                                             BPU  id(0)     HzSElementwiseMul          0.998760           9.6324     int8      
/ReduceMean_13                                      BPU  id(0)     HzSQuantizedReduceMean     0.998329           38.8951    int8      
/gc_pad.13/Pad                                      BPU  id(0)     HzPad                      0.998329           4.77222    int8      
/Slice_24                                           BPU  id(0)     Slice                      0.997473           9.6324     int8      
/Slice_25                                           BPU  id(0)     Slice                      0.996255           9.63273    int8      
/Mul_14                                             BPU  id(0)     HzSElementwiseMul          0.998644           9.6324     int8      
/ReduceMean_14                                      BPU  id(0)     HzSQuantizedReduceMean     0.999202           38.684     int8      
/gc_pad.14/Pad                                      BPU  id(0)     HzPad                      0.999202           4.77222    int8      
/Slice_26                                           BPU  id(0)     Slice                      0.997478           9.6324     int8      
/Slice_27                                           BPU  id(0)     Slice                      0.996294           9.63273    int8      
/Mul_15                                             BPU  id(0)     HzSElementwiseMul          0.998770           9.6324     int8      
/ReduceMean_15                                      BPU  id(0)     HzSQuantizedReduceMean     0.998445           39.0108    int8      
/gc_pad.15/Pad                                      BPU  id(0)     HzPad                      0.998445           4.77222    int8      
/Slice_28                                           BPU  id(0)     Slice                      0.997487           9.6324     int8      
/Slice_29                                           BPU  id(0)     Slice                      0.996309           9.63273    int8      
/Mul_16                                             BPU  id(0)     HzSElementwiseMul          0.998648           9.6324     int8      
/ReduceMean_16                                      BPU  id(0)     HzSQuantizedReduceMean     0.999176           38.7053    int8      
/gc_pad.16/Pad                                      BPU  id(0)     HzPad                      0.999176           4.77222    int8      
/Slice_30                                           BPU  id(0)     Slice                      0.997487           9.6324     int8      
/Slice_31                                           BPU  id(0)     Slice                      0.996349           9.63273    int8      
/Mul_17                                             BPU  id(0)     HzSElementwiseMul          0.998776           9.6324     int8      
/ReduceMean_17                                      BPU  id(0)     HzSQuantizedReduceMean     0.998476           39.0087    int8      
/gc_pad.17/Pad                                      BPU  id(0)     HzPad                      0.998476           4.77222    int8      
/Concat_1                                           BPU  id(0)     Concat                     0.998666           4.77222    int8      
/Mul_18                                             BPU  id(0)     HzSElementwiseMul          0.992203           3.32813    int8      
/ReduceMean_18                                      BPU  id(0)     HzSQuantizedReduceMean     0.999885           11.0743    int8      
/Slice_32                                           BPU  id(0)     Slice                      0.997372           3.32813    int8      
/Slice_33                                           BPU  id(0)     Slice                      0.994305           3.32934    int8      
/Mul_19                                             BPU  id(0)     HzSElementwiseMul          0.991655           3.32813    int8      
/ReduceMean_19                                      BPU  id(0)     HzSQuantizedReduceMean     0.999888           7.47389    int8      
/gc_pad.19/Pad                                      BPU  id(0)     HzPad                      0.999888           1.75165    int8      
/Slice_34                                           BPU  id(0)     Slice                      0.997360           3.32813    int8      
/Slice_35                                           BPU  id(0)     Slice                      0.994270           3.32934    int8      
/Mul_20                                             BPU  id(0)     HzSElementwiseMul          0.991893           3.32813    int8      
/ReduceMean_20                                      BPU  id(0)     HzSQuantizedReduceMean     0.999900           7.54711    int8      
/gc_pad.20/Pad                                      BPU  id(0)     HzPad                      0.999900           1.75165    int8      
/Slice_36                                           BPU  id(0)     Slice                      0.997338           3.32813    int8      
/Slice_37                                           BPU  id(0)     Slice                      0.994257           3.32934    int8      
/Mul_21                                             BPU  id(0)     HzSElementwiseMul          0.991531           3.32813    int8      
/ReduceMean_21                                      BPU  id(0)     HzSQuantizedReduceMean     0.999893           7.48851    int8      
/gc_pad.21/Pad                                      BPU  id(0)     HzPad                      0.999893           1.75165    int8      
/Slice_38                                           BPU  id(0)     Slice                      0.997323           3.32813    int8      
/Slice_39                                           BPU  id(0)     Slice                      0.994267           3.32934    int8      
/Mul_22                                             BPU  id(0)     HzSElementwiseMul          0.991443           3.32813    int8      
/ReduceMean_22                                      BPU  id(0)     HzSQuantizedReduceMean     0.999887           7.50322    int8      
/gc_pad.22/Pad                                      BPU  id(0)     HzPad                      0.999888           1.75165    int8      
/Slice_40                                           BPU  id(0)     Slice                      0.997319           3.32813    int8      
/Slice_41                                           BPU  id(0)     Slice                      0.994275           3.32934    int8      
/Mul_23                                             BPU  id(0)     HzSElementwiseMul          0.991503           3.32813    int8      
/ReduceMean_23                                      BPU  id(0)     HzSQuantizedReduceMean     0.999892           7.48761    int8      
/gc_pad.23/Pad                                      BPU  id(0)     HzPad                      0.999892           1.75165    int8      
/Slice_42                                           BPU  id(0)     Slice                      0.997316           3.32813    int8      
/Slice_43                                           BPU  id(0)     Slice                      0.994284           3.32934    int8      
/Mul_24                                             BPU  id(0)     HzSElementwiseMul          0.991414           3.32813    int8      
/ReduceMean_24                                      BPU  id(0)     HzSQuantizedReduceMean     0.999891           7.50667    int8      
/gc_pad.24/Pad                                      BPU  id(0)     HzPad                      0.999891           1.75165    int8      
/Slice_44                                           BPU  id(0)     Slice                      0.997316           3.32813    int8      
/Slice_45                                           BPU  id(0)     Slice                      0.994295           3.32934    int8      
/Mul_25                                             BPU  id(0)     HzSElementwiseMul          0.991503           3.32813    int8      
/ReduceMean_25                                      BPU  id(0)     HzSQuantizedReduceMean     0.999895           7.48625    int8      
/gc_pad.25/Pad                                      BPU  id(0)     HzPad                      0.999895           1.75165    int8      
/Slice_46                                           BPU  id(0)     Slice                      0.997320           3.32813    int8      
/Slice_47                                           BPU  id(0)     Slice                      0.994309           3.32934    int8      
/Mul_26                                             BPU  id(0)     HzSElementwiseMul          0.991460           3.32813    int8      
/ReduceMean_26                                      BPU  id(0)     HzSQuantizedReduceMean     0.999893           7.50491    int8      
/gc_pad.26/Pad                                      BPU  id(0)     HzPad                      0.999893           1.75165    int8      
/Slice_48                                           BPU  id(0)     Slice                      0.997321           3.32813    int8      
/Slice_49                                           BPU  id(0)     Slice                      0.994320           3.32934    int8      
/Mul_27                                             BPU  id(0)     HzSElementwiseMul          0.991548           3.32813    int8      
/ReduceMean_27                                      BPU  id(0)     HzSQuantizedReduceMean     0.999894           7.48856    int8      
/gc_pad.27/Pad                                      BPU  id(0)     HzPad                      0.999894           1.75165    int8      
/Slice_50                                           BPU  id(0)     Slice                      0.997324           3.32813    int8      
/Slice_51                                           BPU  id(0)     Slice                      0.994327           3.32934    int8      
/Mul_28                                             BPU  id(0)     HzSElementwiseMul          0.991488           3.32813    int8      
/ReduceMean_28                                      BPU  id(0)     HzSQuantizedReduceMean     0.999896           7.5034     int8      
/gc_pad.28/Pad                                      BPU  id(0)     HzPad                      0.999896           1.75165    int8      
/Slice_52                                           BPU  id(0)     Slice                      0.997326           3.32813    int8      
/Slice_53                                           BPU  id(0)     Slice                      0.994334           3.32934    int8      
/Mul_29                                             BPU  id(0)     HzSElementwiseMul          0.991550           3.32813    int8      
/ReduceMean_29                                      BPU  id(0)     HzSQuantizedReduceMean     0.999893           7.48567    int8      
/gc_pad.29/Pad                                      BPU  id(0)     HzPad                      0.999893           1.75165    int8      
/Slice_54                                           BPU  id(0)     Slice                      0.997323           3.32813    int8      
/Slice_55                                           BPU  id(0)     Slice                      0.994341           3.32934    int8      
/Mul_30                                             BPU  id(0)     HzSElementwiseMul          0.991513           3.32813    int8      
/ReduceMean_30                                      BPU  id(0)     HzSQuantizedReduceMean     0.999894           7.50585    int8      
/gc_pad.30/Pad                                      BPU  id(0)     HzPad                      0.999894           1.75165    int8      
/Slice_56                                           BPU  id(0)     Slice                      0.997326           3.32813    int8      
/Slice_57                                           BPU  id(0)     Slice                      0.994351           3.32934    int8      
/Mul_31                                             BPU  id(0)     HzSElementwiseMul          0.991573           3.32813    int8      
/ReduceMean_31                                      BPU  id(0)     HzSQuantizedReduceMean     0.999894           7.48628    int8      
/gc_pad.31/Pad                                      BPU  id(0)     HzPad                      0.999894           1.75165    int8      
/Slice_58                                           BPU  id(0)     Slice                      0.997325           3.32813    int8      
/Slice_59                                           BPU  id(0)     Slice                      0.994360           3.32934    int8      
/Mul_32                                             BPU  id(0)     HzSElementwiseMul          0.991538           3.32813    int8      
/ReduceMean_32                                      BPU  id(0)     HzSQuantizedReduceMean     0.999889           7.50469    int8      
/gc_pad.32/Pad                                      BPU  id(0)     HzPad                      0.999889           1.75165    int8      
/Slice_60                                           BPU  id(0)     Slice                      0.997328           3.32813    int8      
/Slice_61                                           BPU  id(0)     Slice                      0.994369           3.32934    int8      
/Mul_33                                             BPU  id(0)     HzSElementwiseMul          0.991629           3.32813    int8      
/ReduceMean_33                                      BPU  id(0)     HzSQuantizedReduceMean     0.999894           7.48927    int8      
/gc_pad.33/Pad                                      BPU  id(0)     HzPad                      0.999894           1.75165    int8      
/Slice_62                                           BPU  id(0)     Slice                      0.997329           3.32813    int8      
/Slice_63                                           BPU  id(0)     Slice                      0.994379           3.32934    int8      
/Mul_34                                             BPU  id(0)     HzSElementwiseMul          0.991568           3.32813    int8      
/ReduceMean_34                                      BPU  id(0)     HzSQuantizedReduceMean     0.999892           7.5079     int8      
/gc_pad.34/Pad                                      BPU  id(0)     HzPad                      0.999892           1.75165    int8      
/Slice_64                                           BPU  id(0)     Slice                      0.997333           3.32813    int8      
/Slice_65                                           BPU  id(0)     Slice                      0.994392           3.32934    int8      
/Mul_35                                             BPU  id(0)     HzSElementwiseMul          0.991653           3.32813    int8      
/ReduceMean_35                                      BPU  id(0)     HzSQuantizedReduceMean     0.999894           7.48554    int8      
/gc_pad.35/Pad                                      BPU  id(0)     HzPad                      0.999894           1.75165    int8      
/Slice_66                                           BPU  id(0)     Slice                      0.997337           3.32813    int8      
/Slice_67                                           BPU  id(0)     Slice                      0.994402           3.32934    int8      
/Mul_36                                             BPU  id(0)     HzSElementwiseMul          0.991600           3.32813    int8      
/ReduceMean_36                                      BPU  id(0)     HzSQuantizedReduceMean     0.999891           7.51066    int8      
/gc_pad.36/Pad                                      BPU  id(0)     HzPad                      0.999891           1.75165    int8      
/Slice_68                                           BPU  id(0)     Slice                      0.997340           3.32813    int8      
/Slice_69                                           BPU  id(0)     Slice                      0.994413           3.32934    int8      
/Mul_37                                             BPU  id(0)     HzSElementwiseMul          0.991669           3.32813    int8      
/ReduceMean_37                                      BPU  id(0)     HzSQuantizedReduceMean     0.999894           7.48535    int8      
/gc_pad.37/Pad                                      BPU  id(0)     HzPad                      0.999894           1.75165    int8      
/Slice_70                                           BPU  id(0)     Slice                      0.997343           3.32813    int8      
/Slice_71                                           BPU  id(0)     Slice                      0.994423           3.32934    int8      
/Mul_38                                             BPU  id(0)     HzSElementwiseMul          0.991602           3.32813    int8      
/ReduceMean_38                                      BPU  id(0)     HzSQuantizedReduceMean     0.999893           7.51573    int8      
/gc_pad.38/Pad                                      BPU  id(0)     HzPad                      0.999893           1.75165    int8      
/Slice_72                                           BPU  id(0)     Slice                      0.997344           3.32813    int8      
/Slice_73                                           BPU  id(0)     Slice                      0.994435           3.32934    int8      
/Mul_39                                             BPU  id(0)     HzSElementwiseMul          0.991709           3.32813    int8      
/ReduceMean_39                                      BPU  id(0)     HzSQuantizedReduceMean     0.999899           7.48519    int8      
/gc_pad.39/Pad                                      BPU  id(0)     HzPad                      0.999899           1.75165    int8      
/Slice_74                                           BPU  id(0)     Slice                      0.997347           3.32813    int8      
/Slice_75                                           BPU  id(0)     Slice                      0.994445           3.32934    int8      
/Mul_40                                             BPU  id(0)     HzSElementwiseMul          0.991653           3.32813    int8      
/ReduceMean_40                                      BPU  id(0)     HzSQuantizedReduceMean     0.999890           7.50676    int8      
/gc_pad.40/Pad                                      BPU  id(0)     HzPad                      0.999890           1.75165    int8      
/Slice_76                                           BPU  id(0)     Slice                      0.997350           3.32813    int8      
/Slice_77                                           BPU  id(0)     Slice                      0.994456           3.32934    int8      
/Mul_41                                             BPU  id(0)     HzSElementwiseMul          0.991737           3.32813    int8      
/ReduceMean_41                                      BPU  id(0)     HzSQuantizedReduceMean     0.999896           7.4845     int8      
/gc_pad.41/Pad                                      BPU  id(0)     HzPad                      0.999896           1.75165    int8      
/Concat_2                                           BPU  id(0)     Concat                     0.999916           1.75165    int8      
...anches.0/branches.0.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.999378           1.75165    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.0/branches.0.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.997120           5.47949    int8      
...anches.1/branches.1.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996957           4.77222    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.1/branches.1.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998517           3.15341    int8      
...anches.2/branches.2.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996938           3.29574    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.2/branches.2.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.997727           2.00532    int8      
....0.1/fuse_layers.0.1.0/fuse_layers.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.992395           4.11977    int8      
/fusions.0/0_1/0_1.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.994402           5.14349    int8      
....0.2/fuse_layers.0.2.0/fuse_layers.0.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.994952           2.80937    int8      
/fusions.0/0_2/0_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.996453           4.88024    int8      
/fusions.0/0_2/0_2.1/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.997305           2.52679    int8      
/fusions.0/relu/LeakyRelu                           BPU  id(0)     HzLut                      0.997636           6.42436    int8      
.../branches.0.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.1.0.0.0/fuse_layers.1.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996410           6.40381    int8      
....1.2/fuse_layers.1.2.0/fuse_layers.1.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.983497           2.80937    int8      
/fusions.0/1_2/1_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.977646           2.02475    int8      
/fusions.0/relu_1/LeakyRelu                         BPU  id(0)     HzLut                      0.995842           4.71386    int8      
...ranches.0.0/relu/Relu_output_0_calibrated_pad_1  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.0.0/fuse_layers.2.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.994313           6.40381    int8      
...e_layers.2.0.0.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.1.0/fuse_layers.2.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.983660           4.32342    int8      
.../branches.1.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.1.0.0/fuse_layers.2.1.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.997504           4.11977    int8      
/fusions.0/Add_5                                    BPU  id(0)     HzSElementwiseAdd          0.985380           3.35327    int8      
/fusions.0/relu_2/LeakyRelu                         BPU  id(0)     HzLut                      0.996905           3.84166    int8      
...anches.0/branches.0.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995387           6.42436    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.0/branches.0.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.994717           5.51407    int8      
...anches.1/branches.1.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.993776           4.71386    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.1/branches.1.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.997319           4.86614    int8      
...anches.2/branches.2.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998809           3.84166    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.2/branches.2.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998494           2.70305    int8      
....0.1/fuse_layers.0.1.0/fuse_layers.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.994330           4.42103    int8      
/fusions.1/0_1/0_1.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.994178           4.68783    int8      
....0.2/fuse_layers.0.2.0/fuse_layers.0.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995948           4.41464    int8      
/fusions.1/0_2/0_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.997675           2.99934    int8      
/fusions.1/0_2/0_2.1/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.998344           1.12198    int8      
/fusions.1/relu/LeakyRelu                           BPU  id(0)     HzLut                      0.995291           5.43656    int8      
.../branches.0.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.1.0.0.0/fuse_layers.1.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.986172           5.42389    int8      
....1.2/fuse_layers.1.2.0/fuse_layers.1.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998579           4.41464    int8      
/fusions.1/1_2/1_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.999386           8.99588    int8      
/fusions.1/relu_1/LeakyRelu                         BPU  id(0)     HzLut                      0.995107           5.95018    int8      
...ranches.0.0/relu/Relu_output_0_calibrated_pad_1  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.0.0/fuse_layers.2.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.993202           5.42389    int8      
...e_layers.2.0.0.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.1.0/fuse_layers.2.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996787           1.68033    int8      
.../branches.1.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.1.0.0/fuse_layers.2.1.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996818           4.42103    int8      
/fusions.1/Add_5                                    BPU  id(0)     HzSElementwiseAdd          0.998422           2.526      int8      
/fusions.1/relu_2/LeakyRelu                         BPU  id(0)     HzLut                      0.998433           5.34572    int8      
...anches.0/branches.0.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996946           5.43656    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.0/branches.0.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.992978           3.00418    int8      
...anches.1/branches.1.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.985150           4.47593    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.1/branches.1.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.987201           2.75209    int8      
...anches.2/branches.2.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.993784           5.34572    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.2/branches.2.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995517           2.44544    int8      
....0.1/fuse_layers.0.1.0/fuse_layers.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995689           4.45106    int8      
/fusions.2/0_1/0_1.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.997166           3.35783    int8      
....0.2/fuse_layers.0.2.0/fuse_layers.0.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.994766           5.01475    int8      
/fusions.2/0_2/0_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.994946           3.15476    int8      
/fusions.2/0_2/0_2.1/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.994909           1.85688    int8      
/fusions.2/relu/LeakyRelu                           BPU  id(0)     HzLut                      0.995013           5.5477     int8      
.../branches.0.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.1.0.0.0/fuse_layers.1.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.967400           5.42608    int8      
....1.2/fuse_layers.1.2.0/fuse_layers.1.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995486           5.01475    int8      
/fusions.2/1_2/1_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.995548           2.89736    int8      
/fusions.2/relu_1/LeakyRelu                         BPU  id(0)     HzLut                      0.993438           4.15033    int8      
...ranches.0.0/relu/Relu_output_0_calibrated_pad_1  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.0.0/fuse_layers.2.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996682           5.42608    int8      
...e_layers.2.0.0.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.1.0/fuse_layers.2.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998280           0.799723   int8      
.../branches.1.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.1.0.0/fuse_layers.2.1.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.984692           4.45106    int8      
/fusions.2/Add_5                                    BPU  id(0)     HzSElementwiseAdd          0.998032           1.55626    int8      
/fusions.2/relu_2/LeakyRelu                         BPU  id(0)     HzLut                      0.998042           5.75927    int8      
...anches.0/branches.0.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.998371           5.5477     int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.0/branches.0.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996775           1.62252    int8      
...anches.1/branches.1.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.987339           4.15033    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.1/branches.1.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988913           2.12655    int8      
...anches.2/branches.2.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.993203           5.75927    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.2/branches.2.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995954           3.69339    int8      
....0.1/fuse_layers.0.1.0/fuse_layers.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.984570           3.59984    int8      
/fusions.3/0_1/0_1.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.983304           3.39306    int8      
....0.2/fuse_layers.0.2.0/fuse_layers.0.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988981           7.19291    int8      
/fusions.3/0_2/0_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.989669           3.96363    int8      
/fusions.3/0_2/0_2.1/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.989543           2.4629     int8      
/fusions.3/relu/LeakyRelu                           BPU  id(0)     HzLut                      0.994472           5.48462    int8      
.../branches.0.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.1.0.0.0/fuse_layers.1.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995359           5.84995    int8      
....1.2/fuse_layers.1.2.0/fuse_layers.1.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.989367           7.19291    int8      
/fusions.3/1_2/1_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.989660           3.93775    int8      
/fusions.3/relu_1/LeakyRelu                         BPU  id(0)     HzLut                      0.992921           4.40702    int8      
...ranches.0.0/relu/Relu_output_0_calibrated_pad_1  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.0.0/fuse_layers.2.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995583           5.84995    int8      
...e_layers.2.0.0.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.1.0/fuse_layers.2.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.991947           0.667036   int8      
.../branches.1.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.1.0.0/fuse_layers.2.1.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.989568           3.59984    int8      
/fusions.3/Add_5                                    BPU  id(0)     HzSElementwiseAdd          0.996557           2.42239    int8      
/fusions.3/relu_2/LeakyRelu                         BPU  id(0)     HzLut                      0.996590           7.79206    int8      
...anches.0/branches.0.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.995435           5.48462    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.0/branches.0.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.993972           1.55938    int8      
...anches.1/branches.1.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.982973           4.40702    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.1/branches.1.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.962742           2.05217    int8      
...anches.2/branches.2.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996585           7.79206    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.2/branches.2.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.996190           2.41868    int8      
....0.1/fuse_layers.0.1.0/fuse_layers.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.987125           5.02891    int8      
/fusions.4/0_1/0_1.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.992648           3.38798    int8      
....0.2/fuse_layers.0.2.0/fuse_layers.0.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.989584           5.76911    int8      
/fusions.4/0_2/0_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.993587           2.30376    int8      
/fusions.4/0_2/0_2.1/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.994990           0.975474   int8      
/fusions.4/relu/LeakyRelu                           BPU  id(0)     HzLut                      0.991493           4.45399    int8      
.../branches.0.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.1.0.0.0/fuse_layers.1.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988219           4.42671    int8      
....1.2/fuse_layers.1.2.0/fuse_layers.1.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.984052           5.76911    int8      
/fusions.4/1_2/1_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.989098           2.21765    int8      
/fusions.4/relu_1/LeakyRelu                         BPU  id(0)     HzLut                      0.991444           5.67311    int8      
...ranches.0.0/relu/Relu_output_0_calibrated_pad_1  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.0.0/fuse_layers.2.0.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988985           4.42671    int8      
...e_layers.2.0.0.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.0.1.0/fuse_layers.2.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.982927           0.782547   int8      
.../branches.1.0/relu/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
.../fuse_layers.2.1.0.0/fuse_layers.2.1.0.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.972925           5.02891    int8      
/fusions.4/Add_5                                    BPU  id(0)     HzSElementwiseAdd          0.987981           3.50529    int8      
/fusions.4/relu_2/LeakyRelu                         BPU  id(0)     HzLut                      0.989117           4.43591    int8      
...anches.0/branches.0.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.989729           4.45399    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.0/branches.0.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.990945           1.1949     int8      
...anches.1/branches.1.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.988797           5.67311    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.1/branches.1.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.983128           1.83942    int8      
...anches.2/branches.2.0/conv/conv.0/conv.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.967317           4.43591    int8      
...nv/conv.0/conv.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
...anches.2/branches.2.0/conv/conv.1/conv.1.0/Conv  BPU  id(0)     HzSQuantizedConv           0.973870           1.30901    int8      
....0.1/fuse_layers.0.1.0/fuse_layers.0.1.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.991598           6.63811    int8      
/fusions.5/0_1/0_1.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.992393           2.6042     int8      
....0.2/fuse_layers.0.2.0/fuse_layers.0.2.0.0/Conv  BPU  id(0)     HzSQuantizedConv           0.957960           4.27381    int8      
/fusions.5/0_2/0_2.0/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.964929           3.78031    int8      
/fusions.5/0_2/0_2.1/ConvTranspose                  BPU  id(0)     HzSQuantizedConvTranspose  0.968254           2.76771    int8      
/fusions.5/relu/LeakyRelu                           BPU  id(0)     HzLut                      0.986651           4.64225    int8      
/final_conv/final_conv.0/Conv                       BPU  id(0)     HzSQuantizedConv           0.983355           4.64225    int8      
/softmax/Softmax_reducemax_FROM_QUANTIZED_SOFTMAX   BPU  id(0)     HzQuantizedReduceMax       0.798691           7.4715     int16     
/softmax/Softmax_sub_FROM_QUANTIZED_SOFTMAX         BPU  id(0)     HzSElementwiseSub          0.974003           7.4715     int16     
/softmax/Softmax_exp_FROM_QUANTIZED_SOFTMAX         BPU  id(0)     HzLut2Layer                0.881378           11.0903    int16     
/softmax/Softmax_reducesum_FROM_QUANTIZED_SOFTMAX   BPU  id(0)     HzSQuantizedReduceSum      0.954746           1.0        int16     
/softmax/Softmax_reciprocal_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzLut2Layer                0.973493           11.2066    int16     
/softmax/Softmax_mul_FROM_QUANTIZED_SOFTMAX         BPU  id(0)     HzSElementwiseMul          0.916740           1.0        int16     
/disp_conv/Conv                                     BPU  id(0)     HzSQuantizedConv           0.989636           1.0        int16     
/hnet/hnet.0/Conv                                   BPU  id(0)     HzSQuantizedConv           0.996335           3.32813    int8      
/hnet/hnet.2/Relu_output_0_calibrated_pad           BPU  id(0)     HzPad                      --                 --         int8      
/hnet/hnet.3/Conv                                   BPU  id(0)     HzSQuantizedConv           0.996363           1.44052    int8      
/context_net/context_net.0/Conv                     BPU  id(0)     HzSQuantizedConv           0.996553           3.32813    int8      
..._net/context_net.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
/context_zqr/Conv                                   BPU  id(0)     HzSQuantizedConv           0.992193           1.43682    int8      
/Slice_78                                           BPU  id(0)     Slice                      0.995165           0.466052   int8      
/Slice_79                                           BPU  id(0)     Slice                      0.988342           0.466052   int8      
/Slice_80                                           BPU  id(0)     Slice                      0.994877           0.466052   int8      
/corr_lookup/pads.0/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice                                  BPU  id(0)     Slice                      0.994424           --         int8      
/corr_lookup/Mul                                    BPU  id(0)     HzSElementwiseMul          0.992203           3.32813    int8      
/corr_lookup/ReduceMean                             BPU  id(0)     HzSQuantizedReduceMean     0.999885           11.0743    int8      
/corr_lookup/pads.1/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice_1                                BPU  id(0)     Slice                      0.994305           --         int8      
/corr_lookup/Mul_1                                  BPU  id(0)     HzSElementwiseMul          0.991655           3.32813    int8      
/corr_lookup/ReduceMean_1                           BPU  id(0)     HzSQuantizedReduceMean     0.999888           7.47389    int8      
/corr_lookup/pads.2/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice_2                                BPU  id(0)     Slice                      0.994269           --         int8      
/corr_lookup/Mul_2                                  BPU  id(0)     HzSElementwiseMul          0.991893           3.32813    int8      
/corr_lookup/ReduceMean_2                           BPU  id(0)     HzSQuantizedReduceMean     0.999900           7.54711    int8      
/corr_lookup/pads.3/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice_3                                BPU  id(0)     Slice                      0.994255           --         int8      
/corr_lookup/Mul_3                                  BPU  id(0)     HzSElementwiseMul          0.991531           3.32813    int8      
/corr_lookup/ReduceMean_3                           BPU  id(0)     HzSQuantizedReduceMean     0.999893           7.48851    int8      
/corr_lookup/pads.4/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice_4                                BPU  id(0)     Slice                      0.994265           --         int8      
/corr_lookup/Mul_4                                  BPU  id(0)     HzSElementwiseMul          0.991442           3.32813    int8      
/corr_lookup/ReduceMean_4                           BPU  id(0)     HzSQuantizedReduceMean     0.999888           7.50322    int8      
/corr_lookup/pads.5/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice_5                                BPU  id(0)     Slice                      0.994273           --         int8      
/corr_lookup/Mul_5                                  BPU  id(0)     HzSElementwiseMul          0.991505           3.32813    int8      
/corr_lookup/ReduceMean_5                           BPU  id(0)     HzSQuantizedReduceMean     0.999892           7.48761    int8      
/corr_lookup/pads.6/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice_6                                BPU  id(0)     Slice                      0.994281           --         int8      
/corr_lookup/Mul_6                                  BPU  id(0)     HzSElementwiseMul          0.991413           3.32813    int8      
/corr_lookup/ReduceMean_6                           BPU  id(0)     HzSQuantizedReduceMean     0.999891           7.50667    int8      
/corr_lookup/pads.7/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice_7                                BPU  id(0)     Slice                      0.994294           --         int8      
/corr_lookup/Mul_7                                  BPU  id(0)     HzSElementwiseMul          0.991504           3.32813    int8      
/corr_lookup/ReduceMean_7                           BPU  id(0)     HzSQuantizedReduceMean     0.999895           7.48625    int8      
/corr_lookup/pads.8/Pad                             BPU  id(0)     HzPad                      0.994424           3.32934    int8      
/corr_lookup/Slice_8                                BPU  id(0)     Slice                      0.994304           --         int8      
/corr_lookup/Mul_8                                  BPU  id(0)     HzSElementwiseMul          0.991459           3.32813    int8      
/corr_lookup/ReduceMean_8                           BPU  id(0)     HzSQuantizedReduceMean     0.999893           7.50491    int8      
/corr_lookup/Concat                                 BPU  id(0)     Concat                     0.999891           1.75165    int8      
/motion_encoder/convc1/convc1.0/Conv                BPU  id(0)     HzSQuantizedConv           0.999452           1.75165    int8      
...er/convc1/convc1.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
/motion_encoder/convc2/convc2.0/Conv                BPU  id(0)     HzSQuantizedConv           0.999160           3.38275    int8      
variable_2036_Requantize                            BPU  id(0)     HzRequantize               --                 --         int16     
/motion_encoder/convd1/convd1.0/Conv                BPU  id(0)     HzSQuantizedConv           0.948870           0.480154   int8      
...er/convd1/convd1.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
/motion_encoder/convd2/convd2.0/Conv                BPU  id(0)     HzSQuantizedConv           0.904634           7.75248    int8      
/disp_conv/Conv_output_0_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/motion_encoder/Concat                              BPU  id(0)     Concat                     0.884095           6.2499     int8      
/motion_encoder/conv_fuse/conv_fuse.0/Conv          BPU  id(0)     HzSQuantizedConv           0.822548           6.2499     int8      
/motion_encoder/Concat_1                            BPU  id(0)     Concat                     0.846848           8.85097    int8      
/hnet/hnet.5/Relu_output_0_calibrated_Requantize    BPU  id(0)     HzRequantize               --                 --         int8      
/conv_gru/Concat                                    BPU  id(0)     Concat                     0.942216           6.59125    int8      
/conv_gru/convz/Conv                                BPU  id(0)     HzSQuantizedConv           0.978125           8.85097    int8      
/conv_gru/Sigmoid                                   BPU  id(0)     HzLut                      0.996461           5.49897    int8      
/conv_gru/convr/Conv                                BPU  id(0)     HzSQuantizedConv           0.951291           8.85097    int8      
/conv_gru/Sigmoid_1                                 BPU  id(0)     HzLut                      0.943544           5.49897    int8      
/conv_gru/Mul                                       BPU  id(0)     HzSElementwiseMul          0.858949           0.995926   int8      
/conv_gru/Concat_1                                  BPU  id(0)     Concat                     0.848402           8.85097    int8      
/conv_gru/convq/Conv                                BPU  id(0)     HzSQuantizedConv           0.883912           8.85097    int8      
/conv_gru/Sub                                       BPU  id(0)     HzSElementwiseSub          0.942817           1.0        int8      
/conv_gru/Mul_1                                     BPU  id(0)     HzSElementwiseMul          0.924406           0.999259   int8      
/conv_gru/Mul_2                                     BPU  id(0)     HzSElementwiseMul          0.871845           0.995926   int8      
/conv_gru/Add_3                                     BPU  id(0)     HzSElementwiseAdd          0.875392           2.76738    int8      
/disp_head/conv/conv.0/Conv                         BPU  id(0)     HzSQuantizedConv           0.795511           6.67284    int8      
/disp_head/conv/conv.3/Conv                         BPU  id(0)     HzSQuantizedConv           0.722695           4.42568    int8      
/motion_encoder/convd1/convd1.0_1/Conv              BPU  id(0)     HzSQuantizedConv           0.945793           0.471422   int8      
.../convd1/convd1.2_1/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
/motion_encoder/convd2/convd2.0_1/Conv              BPU  id(0)     HzSQuantizedConv           0.872059           7.47107    int8      
...c2/convc2.2/Relu_output_0_calibrated_Requantize  BPU  id(0)     HzRequantize               --                 --         int8      
/Add_1_output_0_calibrated_Requantize               BPU  id(0)     HzRequantize               --                 --         int8      
/motion_encoder_1/Concat                            BPU  id(0)     Concat                     0.857632           6.2499     int8      
/motion_encoder/conv_fuse/conv_fuse.0_1/Conv        BPU  id(0)     HzSQuantizedConv           0.745357           4.75937    int8      
/motion_encoder_1/Concat_1                          BPU  id(0)     Concat                     0.802083           7.28185    int8      
/conv_gru/Add_3_output_0_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/conv_gru_1/Concat                                  BPU  id(0)     Concat                     0.837305           6.67284    int8      
/conv_gru/convz_1/Conv                              BPU  id(0)     HzSQuantizedConv           0.926782           7.28185    int8      
/conv_gru_1/Sigmoid                                 BPU  id(0)     HzLut                      0.990814           5.49897    int8      
/conv_gru/convr_1/Conv                              BPU  id(0)     HzSQuantizedConv           0.852662           7.28185    int8      
/conv_gru_1/Sigmoid_1                               BPU  id(0)     HzLut                      0.923904           5.49897    int8      
/conv_gru_1/Mul                                     BPU  id(0)     HzSElementwiseMul          0.814261           0.995926   int8      
/conv_gru_1/Concat_1                                BPU  id(0)     Concat                     0.803067           7.28185    int8      
/conv_gru/convq_1/Conv                              BPU  id(0)     HzSQuantizedConv           0.874834           7.28185    int8      
/conv_gru_1/Sub                                     BPU  id(0)     HzSElementwiseSub          0.946949           1.0        int8      
/conv_gru_1/Mul_1                                   BPU  id(0)     HzSElementwiseMul          0.815946           0.988774   int8      
/conv_gru_1/Mul_2                                   BPU  id(0)     HzSElementwiseMul          0.800807           0.995926   int8      
/conv_gru_1/Add_3                                   BPU  id(0)     HzSElementwiseAdd          0.855686           3.03908    int8      
/disp_head/conv/conv.0_1/Conv                       BPU  id(0)     HzSQuantizedConv           0.719573           4.136      int8      
/disp_head/conv/conv.3_1/Conv                       BPU  id(0)     HzSQuantizedConv           0.606839           2.34121    int8      
/unfold/conv/Conv                                   BPU  id(0)     HzSQuantizedConv           0.991097           0.453292   int8      
/spx_8/spx_8.0/spx_8.0.0/Conv                       BPU  id(0)     HzSQuantizedConv           0.997400           3.32813    int8      
.../spx_8.0/spx_8.0.2/Relu_output_0_calibrated_pad  BPU  id(0)     HzPad                      --                 --         int8      
/spx_8/spx_8.1/spx_8.1.0/Conv                       BPU  id(0)     HzSQuantizedConv           0.993683           0.92132    int8      
/spx_4/conv1/conv/ConvTranspose                     BPU  id(0)     HzSQuantizedConvTranspose  0.995209           1.06253    int8      
...ck_layers.1/relu/Relu_output_0_calibrated_pad_1  BPU  id(0)     HzPad                      --                 --         int8      
/mod2/mod2.0/Conv                                   BPU  id(0)     HzSQuantizedConv           0.999474           21.1246    int8      
/spx_4/conv2/conv2.0/conv2.0.0/Conv                 BPU  id(0)     HzSQuantizedConv           0.998270           1.2822     int8      
/spx_2/conv1/conv/ConvTranspose                     BPU  id(0)     HzSQuantizedConvTranspose  0.992536           1.84151    int8      
.../head_layer/relu/Relu_output_0_calibrated_pad_1  BPU  id(0)     HzPad                      --                 --         int8      
/mod1/mod1.0/Conv                                   BPU  id(0)     HzSQuantizedConv           0.999313           12.9148    int8      
/spx_2/conv2/conv2.0/conv2.0.0/Conv                 BPU  id(0)     HzSQuantizedConv           0.997985           2.055      int8      
/spx/conv/ConvTranspose                             BPU  id(0)     HzSQuantizedConvTranspose  0.979644           1.86657    int8      
/spx/conv/ConvTranspose_output_0_Requantize         BPU  id(0)     HzRequantize               --                 --         int8      
/spx/act/Relu                                       BPU  id(0)     HzLut2Layer                0.986808           60.5741    int16     
/spx_conv3x3/spx_conv3x3.0/Conv                     BPU  id(0)     HzSQuantizedConv           0.983453           60.5741    int16     
variable_2035_Requantize                            BPU  id(0)     HzRequantize               --                 --         int16     
/softmax2/Softmax                                   CPU  --        HzSoftmax                  0.947051           36.4357    float
2026-05-27 10:59:09,597 INFO The quantized model output:
==============================================================================
Output       Cosine Similarity  L1 Distance  L2 Distance  Chebyshev Distance  
------------------------------------------------------------------------------
pred_unfold  0.991097           0.061552     0.000394     0.202870            
spx_pred     0.947051           0.036238     0.000051     0.622954

,但是我在X5板子上输出视差图的时候,几乎变成了纯色,而且颜色分布也不正确,我在PC上推理是正常的。我尝试将不同的算子分别放在BPU和CPU上,目前也没有解决问题。

下面是我相同模型在PC上运行的结果:

我比较困惑的是,我的量化结果目前表现还不错,为什么量化后的精度这么差。我在网上了解到softmax算子量化精度会很大程度影响最终结果,我这个模型罪魁祸首是不是softmax?但是为什么我把softmax放在cpu,他的结果也没有什么变化呢?对于这个部分我还是有很多疑惑,希望各位大佬帮我分析一下 :folded_hands: :folded_hands: :folded_hands:

建议先使用 Featuremap 进行量化排除前处理的影响

你好,我想问的是,如果将input_type_rt和input_type_train都设置为Featuremap,那么norm_type是不是应该设置为no_preprocess。这样的话校准数据的图像需要进行归一化处理吗?非常感谢 :folded_hands:

  1. 可以在量化过程中将输入类型设置为 NCHW 和 featuremap ,包括 train 和 rt 的两个 type,前处理类型修改为no_preprocess , 这样编译出来的 quantized 模型和 bin 模型所需要的数据,也就是所需要的前处理,和浮点 onnx 完全一致。建议在全 featuremap 的基础上进行准备校准数据,和 bin 模型编译。

    1. 校准数据准备:

    2. 数据源:Dump 送入浮点 ONNX 模型之前 的数据(即已经过 Resize、Normalize、HWC->CHW 等 CPU 预处理后的数据)。

    3. 格式:保存为 .bin 或 .npy。

    4. 优势:此时量化模型(Quantized Model)的输入与浮点 Baseline 完全一致,排除了 BPU 硬件前处理(如 NV12 解码、缩放算法差异)带来的干扰,专注于量化本身的精度损失。

非常感谢,目前我使用featuremap生成的bin文件经过验证,可以正常显示图片,视差图和在PC端的结果相比有些损失,但是整体来说分布一致。然后我使用相同的onnx文件和相同的配置进行nv12的转化,两者的余弦相似度都很高,但是结果仍然不太好。请问目前的话是nv12解码和缩放的问题吗,我目前只会设置校准图像是否归一化处理,请问我还应该在哪些部分排查问题呢?
下面分别是我使用featuremap和nv12的配置文件,配置基本保持一致,不知道是否哪些地方设置的还有问题。

# featuremap

# 模型转化相关的参数

model_parameters:

onnx_model: 'onnx/dif_data.onnx'

march: "bayes-e"

layer_out_dump: False

working_dir: 'model_output_featuremap_dif_data'

output_model_file_prefix: 'model_output_featuremap_dif_data'

# remove_node_type: "Dequantize"

node_info: {                                                                                                                     

#"*": {"ON": "BPU", "InputType": "int16", "OutputType": "int16"},

"/softmax/Softmax": {

"ON": "BPU",

"InputType": "int16",

"OutputType": "int16"

      },

"/softmax2/Softmax": {

"ON": "BPU",

"InputType": "int16",

"OutputType": "int16"

      }

    }






input_parameters:

# 双目 = 两个输入,用分号分隔                                               

input_name: "left;right"                 

input_shape: '1x3x384x640;1x3x384x640'   # 根据实际尺寸改                 

# 运行时输入格式 - NV12

input_type_rt: 'featuremap;featuremap'                                                  

# input_type_rt: 'nv12;nv12'                                                  

input_layout_rt: 'NCHW;NCHW'                                                         

# 训练时格式 - PyTorch一般是RGB + NCHW                                      

# input_type_train: 'yuv444;yuv444'

input_type_train: 'featuremap;featuremap'

input_layout_train: 'NCHW;NCHW'                                             

# 预处理 - 根据训练时normalization设置          

norm_type: "no_preprocess;no_preprocess"                         

# norm_type: 'data_mean_and_scale;data_mean_and_scale'

#ImageNet常用值,换成你训练时实际用的                                      

# mean_value: '128.0 128.0 128.0;128.0 128.0 128.0'

# scale_value: '0.0078125 0.0078125 0.0078125;0.0078125 0.0078125 0.0078125'         







# 模型量化相关参数

calibration_parameters:                                                                                                           

cal_data_dir: './calibration_data_split/left_float32;./calibration_data_split/right_float32'                                                

cal_data_type: 'float32;float32'                                                                                                

calibration_type: 'default'  





# 编译器相关参数

compiler_parameters:

compile_mode: 'latency'

debug: False

optimize_level: 'O3'



# nv12

# 模型转化相关的参数

model_parameters:

onnx_model: 'onnx/dif_data.onnx'

march: "bayes-e"

layer_out_dump: False

working_dir: 'model_output_dif_data_nv12'

output_model_file_prefix: 'model_output_dif_data_nv12'

# remove_node_type: "Dequantize"




node_info: {

"/softmax/Softmax": {

"ON": "BPU",

"InputType": "int16",

"OutputType": "int16"

      },

"/softmax2/Softmax": {

"ON": "BPU",

"InputType": "int16",

"OutputType": "int16"

      },

  }




input_parameters:

# 双目 = 两个输入,用分号分隔                                               

input_name: "left;right"                 

input_shape: '1x3x384x640;1x3x384x640'   # 根据实际尺寸改                 

# 运行时输入格式 - NV12

input_type_rt: 'nv12;nv12'                                                                                                 

input_layout_rt: ''                                                         

# 训练时格式                                    

input_type_train: 'yuv444;yuv444'

input_layout_train: 'NCHW;NCHW'                                             



 # 预处理 - 根据训练时normalization设置                          
 norm_type: 'data_mean_and_scale;data_mean_and_scale'                                                                            
 #ImageNet常用值,换成你训练时实际用的                                      
 mean_value: '128.0 128.0 128.0;128.0 128.0 128.0'
 scale_value: '0.0078125 0.0078125 0.0078125;0.0078125 0.0078125 0.0078125'           







# 模型量化相关参数

calibration_parameters:                                                                                                           

cal_data_dir: './calibration_data_vkitti2_no_norm/left_float32;./calibration_data_vkitti2_no_norm/right_float32'                                                

cal_data_type: 'float32;float32'                                                                                                

calibration_type: 'default'  




#   # 预处理 - 根据训练时normalization设置                              

#   norm_type: 'data_mean_and_scale;data_mean_and_scale'

#   #ImageNet常用值,换成你训练时实际用的                                      

#   mean_value: '128.0 128.0 128.0;128.0 128.0 128.0'

#   scale_value: '0.0078125 0.0078125 0.0078125;0.0078125 0.0078125 0.0078125'         







# # 模型量化相关参数

# calibration_parameters:                                                                                                           

#   cal_data_dir: './calibration_data_vkitti2_no_norm/left_float32;./calibration_data_vkitti2_no_norm//right_float32'                                                

#   cal_data_type: 'float32;float32'                                                                                                

#   calibration_type: 'default'  





# 编译器相关参数

compiler_parameters:

compile_mode: 'latency'

debug: False

optimize_level: 'O3'


下面分别是featuremap和nv12两化模型的可视化图片,我觉得两张图片都可以看到轮廓,但是nv12分布非常窄,目前还不清楚什么原因。

featuremap结果:Final disparity: range=[0.81, 68.12] mean=23.06

nv12结果:Final disparity: range=[15.12, 22.67] mean=17.34

配置BPU进行前处理对校准数据也需要做一些处理,详细可以参考工具链手册或者 rdk_model_zoo/samples/vision/ultralytics_yolo/conversion/mapper.py at rdk_x5 · D-Robotics/rdk_model_zoo · GitHub 这个量化YOLO的脚本

你好,我按照校准数据预处理之后,进行了多次测试,发现不同的校准数据集会导致上面的不同的结果。我使用Sceneflow和vkitti2进行校准数据的生成,发现Sceneflow校准数据会正常而vkitti2校准不正常。。我个人理解,校准数据集的视差范围应该大一些比较好,是按照这个依据选取校准集吗,如果测试真实场景的话,校准数据集是不是真实场景会更好?如果我真实场景的视差范围比较小,又该怎么办?非常感谢之前的解答 :folded_hands:

理论上校准数据集选取 只需要是训练数据集随机抽取 并且能够覆盖训练数据集大部分典型场景即可,可以自由增加数量