qingchuanws/tensorflowlitemicro

Tensorflow Lite Micro is a DL inference framework for microcontrollers based on Google Tensorflow Lite

tensorflowlitemicro's Issues

关于TFlite在K210上支持的问题

您好，看到您在TODO中写到下一步会测试在K210上TFLite的运行情况。想问一下，目前TensorFlow Lite是运行在K210的CPU上还是NPU上？

tensorflow micro

您好， 我对您的github很感兴趣，
最近遇到了写tensorflow的问题，不知道是否可以讨论
请教

我的邮箱是
[email protected]
希望能收到您的邮件。谢谢。

在K210上使用TensorflowLiteMicro报错

之前用的时候是可以的。但是最近一次更新后，在K210上编译TensorFlowLite会报下面的错误：

packages/TensorflowLiteMicro-latest/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc: In function 'TfLiteStatus tflite::ops::micro::conv::EvalQuantizedPerChannel(TfLiteContext*, TfLiteNode*, TfLiteConvParams*, const tflite::ops::micro::conv::OpData&, const TfLiteEvalTensor*, const TfLiteEvalTensor*, const TfLiteEvalTensor*, TfLiteEvalTensor*, TfLiteEvalTensor*)':
packages/TensorflowLiteMicro-latest/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc:338:92: note: #pragma message: CMSIS-NN optimization for conv not available for this target. Using reference kernel.
     "CMSIS-NN optimization for conv not available for this target. Using reference kernel.")
                                                                                            ^
packages/TensorflowLiteMicro-latest/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc:349:44: error: base operand of '->' has non-pointer type 'const tflite::ops::micro::conv::OpData'
   op_params.quantized_activation_min = data->output_activation_min;
                                            ^~
packages/TensorflowLiteMicro-latest/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc:350:44: error: base operand of '->' has non-pointer type 'const tflite::ops::micro::conv::OpData'
   op_params.quantized_activation_max = data->output_activation_max;
                                            ^~
packages/TensorflowLiteMicro-latest/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc:353:22: error: base operand of '->' has non-pointer type 'const tflite::ops::micro::conv::OpData'
       op_params, data->per_channel_output_multiplier,
                      ^~
packages/TensorflowLiteMicro-latest/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc:354:11: error: base operand of '->' has non-pointer type 'const tflite::ops::micro::conv::OpData'
       data->per_channel_output_shift, tflite::micro::GetTensorShape(input),
           ^~
scons: *** [build/packages/TensorflowLiteMicro-latest/tensorflow/lite/micro/kernels/cmsis-nn/conv.o] Error 1
scons: building terminated because of errors.

发现最近更改了micro目录下的SConscript，对于RT_USING_TENSORFLOWLITEMICRO_REFERENCE的情况代码为：

if GetDepend(['PKG_USING_TENSORFLOWLITEMICRO_REFERENCE']):
    src = Glob('kernels/*.cc') + Glob('*.c') + Glob('*.cc') + Glob('benchmarks/*.cc') \
    + Glob('examples/person_detection_experimental/*.cc') \
    + Glob('memory_planner/*.cc') + Glob('testing/*.cc') + Glob('kernels/cmsis-nn/*.cc') \
    + Glob('tools/make/downloads/person_model_int8/*.cc')

    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/*.c')
    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/*.c')
    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/*.c')
    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/*.c')
    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/*.c')
    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/*.c')
    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/*.c')
    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/ReshapeFunctions/*.c')
    src += Glob('tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/*.c')

也就是没有使用CMSIS-NN时也会编译CMSIS-NN库中的东西。同时还删除了kernel目录下原有的REFERENCE情况的算子。想问下这里是否存在问题？