专栏算法工具链模型量化

模型量化

已解决
wangning2021-01-27
287
3

你好,下面是是执行build.sh 时量化模型的输出,请问量化后的节点正常吗?

2021-01-27 20:31:10,419 INFO hbdk-cc parameters:{'optimize-level': 'O2', 'input-source': 'ddr', 'optimize-target': 'fast', 'input-layout': 'NHWC', 'output-layout': 'NHWC'}

[==================================================] 100%

2021-01-27 20:32:24,302 INFO [Wed Jan 27 20:32:24 2021] End to compile the model with march bernoulli2.

2021-01-27 20:32:24,308 INFO The converted model node information:

============================================================================================================================================

Node ON Subgraph Type Cosine Similarity Threshold

--------------------------------------------------------------------------------------------------------------------------------------------

HZ_PREPROCESS_FOR_data BPU id(0) HzSQuantizedPreprocess 0.999980 127.000000

Slice_4 BPU id(0) Slice

Slice_9 BPU id(0) Slice

Slice_14 BPU id(0) Slice

Slice_19 BPU id(0) Slice

Slice_24 BPU id(0) Slice

Slice_29 BPU id(0) Slice

Slice_34 BPU id(0) Slice

Slice_39 BPU id(0) Slice

Concat_40 BPU id(0) Concat 0.999980 1.000133

Conv_41 BPU id(0) HzSQuantizedConv 0.999330 1.000133

LeakyRelu_42 BPU id(0) HzLeakyRelu 0.998485 37.843895

Conv_43 BPU id(0) HzSQuantizedConv 0.989113 37.843895

LeakyRelu_44 BPU id(0) HzLeakyRelu 0.992137 84.632118

Conv_45 BPU id(0) HzSQuantizedConv 0.997573 84.632118

LeakyRelu_46 BPU id(0) HzLeakyRelu 0.998900 12.406853

Conv_47 BPU id(0) HzSQuantizedConv 0.994849 12.406853

LeakyRelu_48 BPU id(0) HzLeakyRelu 0.995358 15.019695

Conv_49 BPU id(0) HzSQuantizedConv 0.995578 15.019695

LeakyRelu_50 BPU id(0) HzLeakyRelu 0.996430 16.856922

UNIT_CONV_FOR_Add_51 BPU id(0) HzSQuantizedConv 0.998112 12.406853

Conv_52 BPU id(0) HzSQuantizedConv 0.997386 24.217426

Conv_53 BPU id(0) HzSQuantizedConv 0.989739 84.632118

Concat_54 BPU id(0) Concat 0.994234 41.526363

UNIT_CONV_FOR_BatchNormalization_55 BPU id(0) HzSQuantizedConv 0.987885 41.526363

LeakyRelu_56 BPU id(0) HzLeakyRelu 0.990439 16.270140

Conv_57 BPU id(0) HzSQuantizedConv 0.985979 16.270140

LeakyRelu_58 BPU id(0) HzLeakyRelu 0.988265 17.780376

Conv_59 BPU id(0) HzSQuantizedConv 0.989903 17.780376

LeakyRelu_60 BPU id(0) HzLeakyRelu 0.992144 7.646636

Conv_61 BPU id(0) HzSQuantizedConv 0.995368 7.646636

LeakyRelu_62 BPU id(0) HzLeakyRelu 0.995689 3.049169

Conv_63 BPU id(0) HzSQuantizedConv 0.989673 3.049169

LeakyRelu_64 BPU id(0) HzLeakyRelu 0.991715 16.207050

Conv_65 BPU id(0) HzSQuantizedConv 0.993278 16.207050

LeakyRelu_66 BPU id(0) HzLeakyRelu 0.985560 4.929499

UNIT_CONV_FOR_Add_67 BPU id(0) HzSQuantizedConv 0.992950 3.049169

Conv_68 BPU id(0) HzSQuantizedConv 0.983751 5.362038

LeakyRelu_69 BPU id(0) HzLeakyRelu 0.976945 10.421040

Conv_70 BPU id(0) HzSQuantizedConv 0.987546 10.421040

LeakyRelu_71 BPU id(0) HzLeakyRelu 0.986219 6.423775

UNIT_CONV_FOR_Add_72 BPU id(0) HzSQuantizedConv 0.992482 5.362038

Conv_73 BPU id(0) HzSQuantizedConv 0.983584 8.265815

LeakyRelu_74 BPU id(0) HzLeakyRelu 0.978504 14.750817

Conv_75 BPU id(0) HzSQuantizedConv 0.982635 14.750817

LeakyRelu_76 BPU id(0) HzLeakyRelu 0.984716 9.860833

UNIT_CONV_FOR_Add_77 BPU id(0) HzSQuantizedConv 0.992692 8.265815

Conv_78 BPU id(0) HzSQuantizedConv 0.993532 12.287702

Conv_79 BPU id(0) HzSQuantizedConv 0.994081 7.646636

Concat_80 BPU id(0) Concat 0.993673 5.919123

UNIT_CONV_FOR_BatchNormalization_81 BPU id(0) HzSQuantizedConv 0.979405 5.919123

LeakyRelu_82 BPU id(0) HzLeakyRelu 0.982828 10.669495

Conv_83 BPU id(0) HzSQuantizedConv 0.988633 10.669495

LeakyRelu_84 BPU id(0) HzLeakyRelu 0.985591 8.058681

Conv_85 BPU id(0) HzSQuantizedConv 0.988790 8.058681

LeakyRelu_86 BPU id(0) HzLeakyRelu 0.980187 10.974434

Conv_87 BPU id(0) HzSQuantizedConv 0.993320 10.974434

LeakyRelu_88 BPU id(0) HzLeakyRelu 0.987604 5.635654

Conv_89 BPU id(0) HzSQuantizedConv 0.985681 5.635654

LeakyRelu_90 BPU id(0) HzLeakyRelu 0.983453 14.573164

Conv_91 BPU id(0) HzSQuantizedConv 0.990699 14.573164

LeakyRelu_92 BPU id(0) HzLeakyRelu 0.974090 6.014710

UNIT_CONV_FOR_Add_93 BPU id(0) HzSQuantizedConv 0.983946 5.635654

Conv_94 BPU id(0) HzSQuantizedConv 0.982494 6.107383

LeakyRelu_95 BPU id(0) HzLeakyRelu 0.969318 10.727690

Conv_96 BPU id(0) HzSQuantizedConv 0.976338 10.727690

LeakyRelu_97 BPU id(0) HzLeakyRelu 0.968251 9.272416

UNIT_CONV_FOR_Add_98 BPU id(0) HzSQuantizedConv 0.979535 6.107383

Conv_99 BPU id(0) HzSQuantizedConv 0.980293 9.771717

LeakyRelu_100 BPU id(0) HzLeakyRelu 0.964420 9.405459

Conv_101 BPU id(0) HzSQuantizedConv 0.966043 9.405459

LeakyRelu_102 BPU id(0) HzLeakyRelu 0.964874 13.235600

UNIT_CONV_FOR_Add_103 BPU id(0) HzSQuantizedConv 0.978884 9.771717

Conv_104 BPU id(0) HzSQuantizedConv 0.987194 15.396114

Conv_105 BPU id(0) HzSQuantizedConv 0.988165 10.974434

Concat_106 BPU id(0) Concat 0.987324 8.208631

UNIT_CONV_FOR_BatchNormalization_107 BPU id(0) HzSQuantizedConv 0.969459 8.208631

LeakyRelu_108 BPU id(0) HzLeakyRelu 0.968911 11.012651

Conv_109 BPU id(0) HzSQuantizedConv 0.980800 11.012651

LeakyRelu_110 BPU id(0) HzLeakyRelu 0.969842 10.415053

Conv_111 BPU id(0) HzSQuantizedConv 0.976177 10.415053

LeakyRelu_112 BPU id(0) HzLeakyRelu 0.963728 11.615922

Conv_113 BPU id(0) HzSQuantizedConv 0.985847 11.615922

LeakyRelu_114 BPU id(0) HzLeakyRelu 0.988865 9.006297

MaxPool_115 BPU id(0) HzQuantizedMaxPool 0.996298 9.006297

MaxPool_116 BPU id(0) HzQuantizedMaxPool 0.997290 9.006297

MaxPool_117 BPU id(0) HzQuantizedMaxPool 0.997630 9.006297

Concat_118 BPU id(0) Concat 0.996423 9.006297

Conv_119 BPU id(0) HzSQuantizedConv 0.973039 9.006297

LeakyRelu_120 BPU id(0) HzLeakyRelu 0.943040 6.750068

Conv_121 BPU id(0) HzSQuantizedConv 0.956309 6.750068

LeakyRelu_122 BPU id(0) HzLeakyRelu 0.943535 10.781240

Conv_123 BPU id(0) HzSQuantizedConv 0.953835 10.781240

LeakyRelu_124 BPU id(0) HzLeakyRelu 0.942783 11.440351

Conv_125 BPU id(0) HzSQuantizedConv 0.966529 11.440351

LeakyRelu_126 BPU id(0) HzLeakyRelu 0.943147 13.059309

Conv_127 BPU id(0) HzSQuantizedConv 0.967990 13.059309

Conv_128 BPU id(0) HzSQuantizedConv 0.963392 6.750068

Concat_129 BPU id(0) Concat 0.965802 3.621874

UNIT_CONV_FOR_BatchNormalization_130 BPU id(0) HzSQuantizedConv 0.947357 3.621874

LeakyRelu_131 BPU id(0) HzLeakyRelu 0.926165 11.522936

Conv_132 BPU id(0) HzSQuantizedConv 0.948074 11.522936

LeakyRelu_133 BPU id(0) HzLeakyRelu 0.926321 13.768301

Conv_134 BPU id(0) HzSQuantizedConv 0.951314 13.768301

LeakyRelu_135 BPU id(0) HzLeakyRelu 0.938877 11.005855

Resize_136 BPU id(0) HzQuantizedResizeUpsample 0.938884 11.005855

UNIT_CONV_FOR_277_0.086660273373127_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv

Concat_137 BPU id(0) Concat 0.954792 11.005855

Conv_138 BPU id(0) HzSQuantizedConv 0.978825 11.005855

LeakyRelu_139 BPU id(0) HzLeakyRelu 0.979678 8.292121

Conv_140 BPU id(0) HzSQuantizedConv 0.969733 8.292121

LeakyRelu_141 BPU id(0) HzLeakyRelu 0.969579 10.753510

Conv_142 BPU id(0) HzSQuantizedConv 0.968012 10.753510

LeakyRelu_143 BPU id(0) HzLeakyRelu 0.962633 10.268023

Conv_144 BPU id(0) HzSQuantizedConv 0.970211 10.268023

Conv_145 BPU id(0) HzSQuantizedConv 0.955068 11.005855

Concat_146 BPU id(0) Concat 0.960288 6.996077

UNIT_CONV_FOR_BatchNormalization_147 BPU id(0) HzSQuantizedConv 0.940865 6.996077

LeakyRelu_148 BPU id(0) HzLeakyRelu 0.945328 11.954016

Conv_149 BPU id(0) HzSQuantizedConv 0.942977 11.954016

LeakyRelu_150 BPU id(0) HzLeakyRelu 0.939499 12.958503

Conv_151 BPU id(0) HzSQuantizedConv 0.954921 12.958503

LeakyRelu_152 BPU id(0) HzLeakyRelu 0.964657 10.759046

Resize_153 BPU id(0) HzQuantizedResizeUpsample 0.964681 10.759046

UNIT_CONV_FOR_251_0.084716893732548_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv

Concat_154 BPU id(0) Concat 0.972211 10.759046

Conv_155 BPU id(0) HzSQuantizedConv 0.987898 10.759046

LeakyRelu_156 BPU id(0) HzLeakyRelu 0.992813 4.302513

Conv_157 BPU id(0) HzSQuantizedConv 0.982609 4.302513

LeakyRelu_158 BPU id(0) HzLeakyRelu 0.988055 8.725763

Conv_159 BPU id(0) HzSQuantizedConv 0.967104 8.725763

LeakyRelu_160 BPU id(0) HzLeakyRelu 0.977691 10.605850

Conv_161 BPU id(0) HzSQuantizedConv 0.977032 10.605850

Conv_162 BPU id(0) HzSQuantizedConv 0.946889 10.759046

Concat_163 BPU id(0) Concat 0.969209 5.671465

UNIT_CONV_FOR_BatchNormalization_164 BPU id(0) HzSQuantizedConv 0.959742 5.671465

LeakyRelu_165 BPU id(0) HzLeakyRelu 0.967933 13.156132

Conv_166 BPU id(0) HzSQuantizedConv 0.942724 13.156132

LeakyRelu_167 BPU id(0) HzLeakyRelu 0.952913 25.990883

Conv_168 BPU id(0) HzSQuantizedConv 0.917437 25.990883

LeakyRelu_169 BPU id(0) HzLeakyRelu 0.927420 10.759046

Concat_170 BPU id(0) Concat 0.949238 10.759046

Conv_171 BPU id(0) HzSQuantizedConv 0.915640 10.759046

LeakyRelu_172 BPU id(0) HzLeakyRelu 0.920992 7.063753

Conv_173 BPU id(0) HzSQuantizedConv 0.915708 7.063753

LeakyRelu_174 BPU id(0) HzLeakyRelu 0.904755 8.051638

Conv_175 BPU id(0) HzSQuantizedConv 0.907962 8.051638

LeakyRelu_176 BPU id(0) HzLeakyRelu 0.920539 9.873170

Conv_177 BPU id(0) HzSQuantizedConv 0.948325 9.873170

Conv_178 BPU id(0) HzSQuantizedConv 0.943825 10.759046

Concat_179 BPU id(0) Concat 0.946474 6.402004

UNIT_CONV_FOR_BatchNormalization_180 BPU id(0) HzSQuantizedConv 0.912207 6.402004

LeakyRelu_181 BPU id(0) HzLeakyRelu 0.927258 9.231904

Conv_182 BPU id(0) HzSQuantizedConv 0.918890 9.231904

LeakyRelu_183 BPU id(0) HzLeakyRelu 0.937252 21.446375

Conv_184 BPU id(0) HzSQuantizedConv 0.918401 21.446375

LeakyRelu_185 BPU id(0) HzLeakyRelu 0.917822 13.046271

UNIT_CONV_FOR_302_0.102726548910141_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv

Concat_186 BPU id(0) Concat 0.925212 13.046271

Conv_187 BPU id(0) HzSQuantizedConv 0.916998 13.046271

LeakyRelu_188 BPU id(0) HzLeakyRelu 0.910209 11.785963

Conv_189 BPU id(0) HzSQuantizedConv 0.901105 11.785963

LeakyRelu_190 BPU id(0) HzLeakyRelu 0.891725 8.725742

Conv_191 BPU id(0) HzSQuantizedConv 0.912393 8.725742

LeakyRelu_192 BPU id(0) HzLeakyRelu 0.905813 10.427575

Conv_193 BPU id(0) HzSQuantizedConv 0.949803 10.427575

Conv_194 BPU id(0) HzSQuantizedConv 0.958761 13.046271

Concat_195 BPU id(0) Concat 0.953877 8.846138

UNIT_CONV_FOR_BatchNormalization_196 BPU id(0) HzSQuantizedConv 0.918535 8.846138

LeakyRelu_197 BPU id(0) HzLeakyRelu 0.927198 11.809894

Conv_198 BPU id(0) HzSQuantizedConv 0.935933 11.809894

LeakyRelu_199 BPU id(0) HzLeakyRelu 0.947075 15.775484

Conv_200 BPU id(0) HzSQuantizedConv 0.993585 25.990883

Conv_202 BPU id(0) HzSQuantizedConv 0.991821 21.446375

Conv_204 BPU id(0) HzSQuantizedConv 0.992369 15.775484

2021-01-27 20:32:24,312 INFO [Wed Jan 27 20:32:24 2021] End to Horizon NN Model Convert.

2021-01-27 20:32:24,321 INFO start convert to *.bin file....

2021-01-27 20:32:24,660 INFO ########################################

2021-01-27 20:32:24,660 INFO ----------- dependency info ------------

2021-01-27 20:32:24,660 INFO hbdk version: 3.13.2

2021-01-27 20:32:24,661 INFO hbdk runtime version: 3.9.8

2021-01-27 20:32:24,661 INFO horizon_nn version: 0.8.1

2021-01-27 20:32:24,661 INFO -------- model parameters info ---------

2021-01-27 20:32:24,661 INFO caffe_model:

2021-01-27 20:32:24,661 INFO prototxt:

2021-01-27 20:32:24,661 INFO onnx_model: /data/wn/x3_tc_1.1.17e/horizon_x3_tc_1.1.17e/samples/04_detection/03_yolov5/mapper/yolov5.onnx

2021-01-27 20:32:24,661 INFO layer_out_dump: False

2021-01-27 20:32:24,661 INFO output_layout: NHWC

2021-01-27 20:32:24,661 INFO -------- input_parameters info ---------

2021-01-27 20:32:24,662 INFO -------- input info : data -------

2021-01-27 20:32:24,662 INFO --input_name : data

2021-01-27 20:32:24,662 INFO --input_type_rt : yuv444_128

2021-01-27 20:32:24,662 INFO --input_type_train : rgbp

2021-01-27 20:32:24,662 INFO --norm_type : data_scale

2021-01-27 20:32:24,662 INFO --input_shape : 1x3x672x672

2021-01-27 20:32:24,662 INFO ----------------------------------

2021-01-27 20:32:24,662 INFO -------- calibration parameters info ---------

2021-01-27 20:32:24,663 INFO preprocess_on: False

2021-01-27 20:32:24,663 INFO calibration_type: max

2021-01-27 20:32:24,663 INFO promoter_level: -1

2021-01-27 20:32:24,663 INFO per_channel: False

2021-01-27 20:32:24,663 INFO max_percentile: 1.0

2021-01-27 20:32:24,663 INFO ------------ compiler_parameters info -------------

2021-01-27 20:32:24,663 INFO compile_mode: latency

2021-01-27 20:32:24,663 INFO debug: False

2021-01-27 20:32:24,664 INFO optimize_level: O2

2021-01-27 20:32:24,664 INFO input_source: {'data': 'ddr'}

2021-01-27 20:32:24,664 INFO ########################################

2021-01-27 20:32:24,680 INFO Convert to runtime bin file sucessfully!

2021-01-27 20:32:24,680 INFO End Model Convert

算法工具链
评论1
0/1000
  • admin
    Lv.1
    你好,量化转换日志里面的节点的量化和浮点之间的相似度是基于第一张校准图的对比结果,其中尾部节点的相似度如果大于95%,即代表量化可能效果还不错,如果相似度低于80%,即代表量化较差。

    但是最终的量化精度建议采用后处理的代码进行算法指标验证,此处的相似度大小和最终算法精度之间没有必然线性关系。

    2021-01-28
    0
    2
    • Huanyong_Ji回复admin:

      Hi,

      Thanks for your reply!My question is:

      1,Quantization method is just Nvidia's method or other?

      2,Can i get a instruction about which type of DNN model can be accelerated by Quantization.

      Thanks!

      2021-02-25
      0
    • ggqmchy回复admin:

      你好,能不能回复一下我发的帖子

      2021-03-04
      0