tensorflow - 无论提供任何图像，TFLite 都会得到相同的结果
问题描述
我尝试在 Android TFLite 中部署它，但无论我提供任何图像，解释器总是向我发送相同的结果
我使用 tf.keras 构建了一个 FaceNet 模型并尝试将其部署在 Android TFLite 中，转换（从 tf.keras 到 tflite）的处理变得平滑。
我尝试使用 tflite python api 进行推理，它有效。但是当我尝试在 Android 中部署它时，无论我提供任何图像，解释器总是给我发送相同的结果。
Android中有推理代码。
// Init
Activity activity = (Activity) othersMap.get("activity");
modelFile = loadModelFile(activity);
tfLite = new Interpreter(modelFile, tfliteOptions);
useCPU();
imgData = ByteBuffer.allocateDirect(
DIM_BATCH_SIZE * DIM_IMG_SIZE_X * DIM_IMG_SIZE_Y * DIM_PIXEL_SIZE * 4);
imgData.order(ByteOrder.nativeOrder());
embeddingArray = new float[1][128];
// DetectImage
Bitmap bmp = Bitmap.createBitmap(src.width(), src.height(),
Bitmap.Config.ARGB_8888);
Utils.matToBitmap(src, bmp);
convertBitmapToByteBuffer(bmp);
tfLite.resizeInput(0, new int[]{1, 96, 96, 3});
tfLite.run(imgData, embeddingArray);
// Inference
float[][] embededFace = detector.detectImage(faceMat);
float[][] embededZN = detector.detectImage(znMat);
Log.e(TAG, "--------------");
Log.e(TAG, Arrays.deepToString(embededFace));
Log.e(TAG, Arrays.deepToString(embededZN));
Log.e(TAG, "--------------");
并且python中有推理代码，它可以工作。
img = cv.imread('zn.jpg')
img = cv.cvtColor(img, cv.COLOR_BGR2RGB)
img = cv.resize(img, (96, 96), interpolation=cv.INTER_CUBIC)
img = np.around(img/255.0, decimals=12)
x_train = np.array([img]).astype(np.float32)
input_shape = input_details[0]['shape']
input_data = x_train
interpreter.set_tensor(input_details[0]['index'], input_data)
interpreter.invoke()
output_data = interpreter.get_tensor(output_details[0]['index'])
print(output_data)
使用tflite android inference api有结果（使用zn.jpg和jzx.jpg）
2019-03-26 15:01:46.346 11635-11823/org.blackwalnutlabs.angel.facedetection E/MainActivity: --------------
2019-03-26 15:01:46.347 11635-11823/org.blackwalnutlabs.angel.facedetection E/MainActivity: [[0.13254969, 0.06506284, -0.12454985, -0.033946857, 0.13087335, 0.24849732, 0.13556245, -0.04702322, -0.1379856, -0.007572789, 0.019669764, 0.012697782, 0.13016781, -0.08282088, 0.07000842, -0.13634653, -0.033497285, 0.06713182, -0.056609895, 0.13066444, 0.031595007, 0.010904907, 0.06502004, 0.055475105, -0.042641174, -0.11620673, -0.13694304, -0.1973542, 0.023469506, 0.105563894, -0.018650757, 0.03307536, -0.098828435, 0.08058539, 0.08887855, 0.0022337863, 0.0055133225, 0.05411017, -0.044450507, -0.082797125, 0.07151169, -0.068259895, -0.09238541, 0.051023263, -0.26672673, 0.062153224, 0.089065045, 0.1272812, -0.15051614, 0.093154915, -0.04973236, -0.055004556, 0.028005369, 9.352369E-4, 0.021919679, -0.019822104, -0.02774606, 0.12849025, -0.051899977, -0.08593918, -0.08694194, 0.10073202, 0.20436272, -0.19222663, 0.084071614, -0.011063285, 0.037152126, 0.03323839, -0.13506632, -0.026216896, 0.012508951, 0.06862544, -0.023205, 0.013072543, -0.031873662, 0.025892422, -0.061958347, -0.06703131, 0.10233366, 0.06788932, -0.02056665, 0.0018195248, -0.07498662, -0.05542643, -0.01115404, 0.009607625, 0.029644933, -0.020729594, 0.0023015668, 0.09378674, 0.18983097, -0.11676965, -0.01661709, -0.019705301, -0.11590064, -0.13303259, 0.0070525194, 0.056587018, 0.02837071, -0.045306873, 0.105889864, 0.0897621, 0.0022198055, 0.043863844, -0.16235375, 0.06855683, -0.04233673, 0.10779853, -0.019658986, 0.096174344, 0.14861028, 0.09041121, -0.03691038, -0.040842745, -0.044187892, 0.06624289, -0.13923372, -0.038001977, -0.018037193, 0.114912674, -7.4458803E-4, -0.122649185, -0.010618781, 0.11189667, -0.03145326, 0.036915697, -0.0388509, -0.096018195]]
2019-03-26 15:01:46.347 11635-11823/org.blackwalnutlabs.angel.facedetection E/MainActivity: [[0.13254969, 0.06506284, -0.12454985, -0.033946857, 0.13087335, 0.24849732, 0.13556245, -0.04702322, -0.1379856, -0.007572789, 0.019669764, 0.012697782, 0.13016781, -0.08282088, 0.07000842, -0.13634653, -0.033497285, 0.06713182, -0.056609895, 0.13066444, 0.031595007, 0.010904907, 0.06502004, 0.055475105, -0.042641174, -0.11620673, -0.13694304, -0.1973542, 0.023469506, 0.105563894, -0.018650757, 0.03307536, -0.098828435, 0.08058539, 0.08887855, 0.0022337863, 0.0055133225, 0.05411017, -0.044450507, -0.082797125, 0.07151169, -0.068259895, -0.09238541, 0.051023263, -0.26672673, 0.062153224, 0.089065045, 0.1272812, -0.15051614, 0.093154915, -0.04973236, -0.055004556, 0.028005369, 9.352369E-4, 0.021919679, -0.019822104, -0.02774606, 0.12849025, -0.051899977, -0.08593918, -0.08694194, 0.10073202, 0.20436272, -0.19222663, 0.084071614, -0.011063285, 0.037152126, 0.03323839, -0.13506632, -0.026216896, 0.012508951, 0.06862544, -0.023205, 0.013072543, -0.031873662, 0.025892422, -0.061958347, -0.06703131, 0.10233366, 0.06788932, -0.02056665, 0.0018195248, -0.07498662, -0.05542643, -0.01115404, 0.009607625, 0.029644933, -0.020729594, 0.0023015668, 0.09378674, 0.18983097, -0.11676965, -0.01661709, -0.019705301, -0.11590064, -0.13303259, 0.0070525194, 0.056587018, 0.02837071, -0.045306873, 0.105889864, 0.0897621, 0.0022198055, 0.043863844, -0.16235375, 0.06855683, -0.04233673, 0.10779853, -0.019658986, 0.096174344, 0.14861028, 0.09041121, -0.03691038, -0.040842745, -0.044187892, 0.06624289, -0.13923372, -0.038001977, -0.018037193, 0.114912674, -7.4458803E-4, -0.122649185, -0.010618781, 0.11189667, -0.03145326, 0.036915697, -0.0388509, -0.096018195]]
2019-03-26 15:01:46.347 11635-11823/org.blackwalnutlabs.angel.facedetection E/MainActivity: --------------
使用tflite python inference api有结果（使用zn.jpg和jzx.jpg）
[[ 0.11388897 0.0725093 -0.02605356 -0.01048064 0.13486011 0.2478467
0.14440472 -0.08799922 -0.09042776 -0.01699087 -0.01304216 -0.00571998
0.08452216 0.0200549 0.15689915 -0.19767451 0.01055939 0.04823723
-0.0236162 0.03551559 -0.03713445 0.07919028 0.00384108 0.07863859
0.01265915 -0.16044462 -0.1557556 -0.08875822 -0.00050363 0.09660906
0.04116863 0.02890096 -0.13043067 0.05401503 0.11558161 0.04363808
0.01493979 -0.02636928 -0.0312182 -0.06017392 0.11421517 -0.02729253
-0.0982411 -0.03221652 -0.22682378 0.09350454 0.00194081 0.1348505
-0.08131604 0.01773626 -0.02530179 -0.00231936 0.09337027 -0.03695107
0.00894977 0.01499904 -0.08354656 0.15669909 -0.05902651 -0.01955242
-0.09552951 0.14642344 0.10130911 -0.20985658 0.07511081 0.05583977
0.08974755 0.03895326 -0.17041707 0.01303602 0.07159603 0.08509548
-0.01078814 -0.01519361 -0.02482661 0.07998506 -0.06343063 -0.12341296
0.10723062 0.0521314 0.00146189 0.02375979 -0.02791258 -0.04265117
0.02753739 -0.0534174 0.04096378 -0.04856863 -0.07195448 0.1120934
0.1552053 -0.09029299 0.07750953 0.00552343 -0.13981637 -0.13607475
0.0221287 -0.03371112 0.04638224 -0.01829939 0.11707122 0.12880346
-0.00621982 -0.00639387 -0.17476223 0.07756975 -0.02970012 0.00726423
-0.0368601 0.13965365 0.17757982 0.02349343 -0.04115034 -0.03476189
-0.06320279 0.04965549 -0.14846669 0.01701067 -0.05107493 0.10043616
0.02471918 -0.14686176 -0.05839831 0.07405312 0.02949061 0.07611959
-0.0424039 -0.09250365]]
[[ 0.07983646 0.18764469 -0.1073711 -0.05049998 0.00499822 0.12150028
0.08042646 -0.01210295 -0.05991281 -0.04436503 0.05246051 -0.01235358
0.08685836 -0.13629174 -0.02383062 -0.06484984 -0.04617283 -0.04979375
-0.10104349 0.14407381 0.03841608 -0.02601014 -0.02829801 0.02454884
-0.04299033 -0.00861871 -0.03450659 -0.12293942 0.09096454 0.08878913
-0.03709067 0.05936524 -0.01129172 0.09354664 0.02958448 -0.03306518
0.10459264 0.15117857 -0.02248681 -0.08084158 0.06213554 -0.16485777
0.03646875 0.11903016 -0.11887202 0.1105961 0.177684 0.04797998
-0.24575731 0.05616427 -0.06487465 0.02870647 -0.03513253 0.04599205
0.09469782 -0.02022276 -0.11111834 0.03512118 -0.04118481 -0.17735724
-0.08443499 -0.04586984 0.1915695 -0.09305634 0.01631605 -0.00732839
0.05056106 -0.02879676 -0.07254668 -0.05057898 -0.05678427 -0.06486657
-0.05519347 0.04775465 0.09962408 -0.04996048 -0.02798802 0.07482754
0.05294487 0.09077998 -0.02322999 0.03212555 -0.07331605 -0.06148532
0.02072088 0.03205773 -0.01099079 -0.01667416 0.06782145 0.12291647
0.13593872 -0.18584044 -0.08280354 -0.04093247 -0.08853628 -0.08608698
0.05808447 0.05453911 0.05205403 -0.11206362 -0.01309075 0.06881329
0.01091391 0.12481726 -0.19396386 0.06527282 -0.06483501 0.23652825
0.00085458 -0.05038064 0.10557921 0.1129869 -0.16408154 -0.00994272
-0.00418024 -0.03208672 -0.13337857 -0.15857975 0.06981471 0.08215429
0.03227568 0.02853267 0.06400527 0.15060182 0.0306827 -0.04676049
0.0670058 -0.10612144]]
解决方案
推荐阅读
- android - 如何在 Android 的非活动类中编写自定义警报？
- oracle - 使用 UTL_FILE ORACLE PACKAGE 导入 xls、xlsx 文件
- amazon-web-services - 如何从 AWS DMS 中查找 Secret？
- javascript - 如何在javascript中创建一个没有新声明的点函数？
- google-sheets - Google 表格 - 如果名称出现在单独的列表中，则使用条件格式点亮列表中的名称
- r - 从 UTC 到本地日期时间的 SQL Server 时区（通过 dbplyr）
- android - AOSP 的安全测试套件 (STS) 适用性
- pine-script - 我的 pinescript 代码中回测数据为 0 的原因
- angular - 更新 Angular 本地库项目
- c++ - 如何用for的环把字母组合成一个词？