深度學習模型檔案mnn量化實踐
阿新 • • 發佈：2021-01-08
轉化成mnn模型雖然可以進行推理
不過模型檔案可能較大或者執行較慢的情況
特別是在移動裝置等邊緣裝置上，算力和儲存空間受限
因此壓縮模型是一個急需的工作
mnn自帶了量化工具，環境安裝很簡單，這文章編譯就可以使用量化了
mnn模型檔案是使用的是之前的文章訓練並轉化的mnn檔案
在使用之前需要新建一個json檔案，裡面配置好內容
preprocessConfig.json
{ "format":"GRAY", "mean":[ 127.5 ], "normal":[ 0.00784314 ], "width":28, "height":28, "path":"FashionMNIST", "used_image_num":50, "feature_quantize_method":"KL", "weight_quantize_method":"MAX_ABS" }
需要說明的是因為之前那篇檔案訓練的是單通道的圖片，因此json檔案裡面format寫的GRAY，實際上可以選擇的有："RGB", "BGR", "RGBA", "GRAY"，需要根據自己模型情況進行選擇
具體可以看一下文件https://www.yuque.com/mnn/cn/tool_quantize
準備工作都做好了，現在只需要量化即可
/opt/MNN/build/quantized.out FashionMNIST.mnn quan.mnn preprocessConfig.json量化的結果：
[15:54:55] /opt/MNN/tools/quantization/quantized.cpp:21: >>> modelFile: FashionMNIST.mnn [15:54:55] /opt/MNN/tools/quantization/quantized.cpp:22: >>> preTreatConfig: preprocessCo nfig.json [15:54:55] /opt/MNN/tools/quantization/quantized.cpp:23: >>> dstFile: quan.mnn [15:54:55] /opt/MNN/tools/quantization/quantized.cpp:50: Calibrate the feature and quanti ze model... [15:54:55] /opt/MNN/tools/quantization/calibration.cpp:121: Use feature quantization meth od: KL [15:54:55] /opt/MNN/tools/quantization/calibration.cpp:122: Use weight quantization metho d: MAX_ABS [15:54:55] /opt/MNN/tools/quantization/Helper.cpp:100: used image num: 50 ComputeFeatureRange: 100.00 % CollectFeatureDistribution: 100.00 % [15:54:55] /opt/MNN/tools/quantization/quantized.cpp:54: Quantize model done!
389K的檔案變成了108K
現在需要測試一下量化和未量化之前的執行速度和執行結果
import time import MNN import numpy as np if __name__ == '__main__': x=np.ones([1, 1, 28, 28]).astype(np.float32) #quan mnn start=time.time() interpreter = MNN.Interpreter("quan.mnn") print("quan mnn load") mnn_session = interpreter.createSession() input_tensor = interpreter.getSessionInput(mnn_session) tmp_input = MNN.Tensor((1, 1, 28, 28),\ MNN.Halide_Type_Float, x[0], MNN.Tensor_DimensionType_Tensorflow) interpreter.runSession(mnn_session) output_tensor = interpreter.getSessionOutput(mnn_session,'output') output_data=np.array(output_tensor.getData()) print('quan mnn result is:',output_data) print('quan mnn run time is ',time.time()-start)
執行結果：
quan mnn load
quan mnn result is: [ 0.5922392 -0.40196353 0.32656723 0.13848761 0.01854512 -1.11787963
0.99948055 -0.32638997 0.92734373 -0.93912888]
quan mnn run time is 0.0015997886657714844
和之前的執行結果有一定差距，量化必然會帶來精度的損失，需要重點注意
這裡貼出來之前的結果
再使用mnn自帶的時間測試工具測試一下
/opt/MNN/build/timeProfile.out quan.mnn 10 0執行結果：
Use extra forward type: 0
Open Model quan.mnn
Sort by node name !
Node Name Op Type Avg(ms) % Flops Rate
11 ConvInt8 0.572700 64.814415 5.521849
13 Pooling 0.015200 1.720236 0.585116
13___FloatToInt8___0 FloatToInt8 0.011900 1.346764 0.146279
14 ConvInt8 0.066800 7.559983 44.174789
16 Pooling 0.010500 1.188321 0.292558
16___FloatToInt8___0 FloatToInt8 0.007600 0.860118 0.073140
17 ConvInt8 0.048200 5.454958 44.174789
19 Pooling 0.008000 0.905387 0.107470
23 BinaryOp 0.007800 0.882753 0.005971
MatMul15 ConvInt8 0.004400 0.497963 3.606106
MatMul19 ConvInt8 0.001000 0.113173 0.062606
Raster12 Raster 0.014000 1.584428 0.026868
Raster16 Raster 0.009200 1.041195 0.005971
Raster18 Raster 0.006700 0.758262 0.005971
Raster22 Raster 0.006900 0.780897 0.000466
Reshape14___tr4MatMul15___FloatToInt8___0 FloatToInt8 0.009000 1.018561 0.026868
Reshape19___tr4MatMul19___FloatToInt8___0 FloatToInt8 0.006600 0.746944 0.005971
___Int8ToFloat___For_130 Int8ToFloat 0.021800 2.467180 0.585116
___Int8ToFloat___For_160 Int8ToFloat 0.007400 0.837483 0.292558
___Int8ToFloat___For_190 Int8ToFloat 0.008200 0.928022 0.146279
___Int8ToFloat___For_MatMul15___tr4Reshape160 Int8ToFloat 0.019600 2.218199 0.005971
___Int8ToFloat___For_MatMul19___tr4Reshape210 Int8ToFloat 0.010800 1.222273 0.000560
input___FloatToInt8___0 FloatToInt8 0.004100 0.464011 0.146279
output BinaryOp 0.005200 0.588502 0.000466
Sort by time cost !
Node Type Avg(ms) % Called times Flops Rate
BinaryOp 0.013000 1.471254 2.000000 0.006437
Pooling 0.033700 3.813944 3.000000 0.985145
Raster 0.036800 4.164783 4.000000 0.039275
FloatToInt8 0.039200 4.436398 5.000000 0.398536
Int8ToFloat 0.067800 7.673157 5.000000 1.030484
ConvInt8 0.693100 78.440491 5.000000 97.540123
total time : 0.883600 ms, total mflops : 2.044533
main, 113, cost time: 13.603001 ms
未量化的時間測試
/opt/MNN/build/timeProfile.out FashionMNIST.mnn 10 0執行結果：
Use extra forward type: 0
Open Model FashionMNIST.mnn
Sort by node name !
Node Name Op Type Avg(ms) % Flops Rate
11 Convolution 0.485900 54.638489 5.601901
13 Pooling 0.017400 1.956595 0.593599
14 Convolution 0.089500 10.064097 44.815208
16 Pooling 0.012000 1.349376 0.296799
17 Convolution 0.048400 5.442484 44.815208
19 Pooling 0.153200 17.227037 0.109028
23 BinaryOp 0.006300 0.708422 0.006057
MatMul15 Convolution 0.025700 2.889914 3.658385
MatMul19 Convolution 0.009000 1.012032 0.063514
Raster10 Raster 0.006400 0.719667 0.006057
Raster12 Raster 0.005400 0.607219 0.000473
Raster6 Raster 0.014200 1.596762 0.027257
Raster8 Raster 0.009700 1.090746 0.006057
output BinaryOp 0.006200 0.697178 0.000473
Sort by time cost !
Node Type Avg(ms) % Called times Flops Rate
BinaryOp 0.012500 1.405600 2.000000 0.006530
Raster 0.035700 4.014394 4.000000 0.039845
Pooling 0.182600 20.533010 3.000000 0.999427
Convolution 0.658500 74.047020 5.000000 98.954201
total time : 0.889300 ms, total mflops : 2.015316
main, 113, cost time: 12.360001 ms