我们正在尝试在android应用中进行对象检测。为此,我们使用了ssdlite_mobilenet_v2_coco预训练网络并继续训练我们自己的数据集。
要进行培训,我们使用了Ubuntu 18.04和TensorFlow 1.11.0 + GPU支持。我们使用以下脚本创建了tflite模型:
python3 ~/tensorflow/models/research/object_detection/export_tflite_ssd_graph.py \
--pipeline_config_path input/ssdlite_mobilenet_v2_coco/pipeline.config \
--trained_checkpoint_prefix input/ssdlite_mobilenet_v2_coco/model.ckpt-381700 \
--output_directory output/ \
--add_postprocessing_op=true
tflite_convert \
--output_file=output/ssdlite_mobilenet_v2_coco.tflite \
--graph_def_file=input/ssdlite_mobilenet_v2_coco.pb \
--input_arrays=FLOAT \
--output_arrays=concat,concat_1 \
--input_shape=1,300,300,3
此应用程序的基本操作是,它会录制预录的视频,然后使用FFmpegMediaMetadataRetriever逐帧对其进行解码,然后将位图传递到tflite中以检测其中的对象。该应用程序是使用gradle构建的,我们使用的是“ org.tensorflow:tensorflow-lite:1.12.0”,但从1.11开始我们基本上会遇到相同的错误。
我们将位图缩小到300x300,并将其从ARGB转换为3个浮动通道,然后按以下方式调用tflite:
Log.v(TAG, "Feeding TFLite")
outputLocations = Array(1) { Array(NUM_DETECTIONS) { FloatArray(4) } }
outputClasses = Array(1) { FloatArray(NUM_DETECTIONS) }
outputScores = Array(1) {FloatArray(NUM_DETECTIONS)}
numDetections = FloatArray(1)
val inputArray = arrayOf<Any>(imgData!!)
val outputMap = HashMap<Int, Any>()
outputMap.put(0, outputLocations!!)
outputMap.put(1, outputClasses!!)
outputMap.put(2, outputScores!!)
outputMap.put(3, numDetections!!)
Log.v(TAG, "Running TFLite")
tflite!!.runForMultipleInputsOutputs(inputArray, outputMap)
Log.v(TAG, "Returning from TFLite")
val recognitions = ArrayList<Recognition>(NUM_DETECTIONS)
我们得到的错误是:
2019-02-28 11:52:33.486 26807-26879/com.package.xxxxxx A/libc: Fatal signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0x0 in tid 26879 (.xxxxxx), pid 26807 (.xxxxxx)
2019-02-28 11:52:33.600 26890-26890/? A/DEBUG: *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***
2019-02-28 11:52:33.601 26890-26890/? A/DEBUG: Build fingerprint: 'google/sdk_gphone_x86/generic_x86:9/PSR1.180720.075/5124027:userdebug/dev-keys'
2019-02-28 11:52:33.601 26890-26890/? A/DEBUG: Revision: '0'
2019-02-28 11:52:33.601 26890-26890/? A/DEBUG: ABI: 'x86'
2019-02-28 11:52:33.604 26890-26890/? A/DEBUG: pid: 26807, tid: 26879, name: .xxxxxx >>> com.package.xxxxxx <<<
2019-02-28 11:52:33.604 26890-26890/? A/DEBUG: signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0x0
2019-02-28 11:52:33.604 26890-26890/? A/DEBUG: Cause: null pointer dereference
2019-02-28 11:52:33.604 26890-26890/? A/DEBUG: eax 00000000 ebx 00000000 ecx 00000000 edx 00000000
2019-02-28 11:52:33.604 26890-26890/? A/DEBUG: edi c75094a8 esi 00000000
2019-02-28 11:52:33.604 26890-26890/? A/DEBUG: ebp c75090f8 esp c7509070 eip c757c77b
2019-02-28 11:52:33.606 26890-26890/? A/DEBUG: backtrace:
2019-02-28 11:52:33.606 26890-26890/? A/DEBUG: #00 pc 0007277b /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.606 26890-26890/? A/DEBUG: #01 pc 00074fe0 /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.606 26890-26890/? A/DEBUG: #02 pc 0007590d /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.606 26890-26890/? A/DEBUG: #03 pc 000755b0 /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.606 26890-26890/? A/DEBUG: #04 pc 00076322 /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.606 26890-26890/? A/DEBUG: #05 pc 0013389c /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.607 26890-26890/? A/DEBUG: #06 pc 00132fa7 /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.607 26890-26890/? A/DEBUG: #07 pc 00132e37 /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.607 26890-26890/? A/DEBUG: #08 pc 0016550e /data/app/com.package.xxxxxx-yGil4a0ylttBBhPIo26SWA==/lib/x86/libtensorflowlite_jni.so
2019-02-28 11:52:33.607 26890-26890/? A/DEBUG: #09 pc 0008f065 /system/lib/libc.so (__pthread_start(void*)+53)
2019-02-28 11:52:33.607 26890-26890/? A/DEBUG: #10 pc 0002485b /system/lib/libc.so (__start_thread+75)
2019-02-28 11:52:34.197 1761-1761/? E//system/bin/tombstoned: Tombstone written to: /data/tombstones/tombstone_35
如您所见,NPE发生在libtensorflow的深处,并且我们基本上没有办法解决该问题,因此可以提供任何帮助。它同时在物理设备和android沙箱(API 28)上发生
我们以this为起点,并且Tensorflow tflite演示也是从tensorflow仓库中形成的。
答案 0 :(得分:0)
事实证明,tflite_convert脚本中的输入数组和输出数组是错误的。
我们设法使用以下命令生成的tflite图成功(没有NullPointer)调用tflite:
tflite_convert \
--graph_def_file=tflite_graph.pb \
--output_file=output.tflite \
--input_shapes=1,300,300,3 \
--std_dev_values=128 \
--mean_values=128 \
--input_arrays=normalized_input_image_tensor \
--output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3' \
--allow_custom_ops