구차니의 잡동사니 모음

[링크 : http://theeye.pe.kr/archives/1980]

[링크 : https://git-scm.com/book/ko/v2/Git-브랜치-Rebase-하기]

[링크 : https://jusths.tistory.com/60]

segmentation fault가 떠서 어디서 죽나 찾는데 꽤나 고생했네 -_-

(빌드가 느리면 역시 스트레스..)

그런데 도무지 allocateTensors() 라는 함수가 어떤 역활을 하는지 모르겠네..

한번 할당하고 또 할당하면 오류나는 느낌..

[링크 : https://www.tensorflow.org/lite/guide/inference]

[링크 : https://www.tensorflow.org/lite/api_docs/cc/class/tflite/interpreter#allocatetensors]

여기 소스를 많이 참고했는데.. tensorflow의 label_image쪽이랑은 또 많이 달라서

어느 가이드를 따라가야 하나 고민..

[링크 : https://github.com/Qengineering/TensorFlow_Lite_SSD_RPi_64-bits/blob/master/MobileNetV1.cpp]

label_image 예제를 보면 get_top_n 에서

(prediction[i] + 128) / 256.0 나 prediction[i] / 255.0 

같은 quantization 범위에 맞춘 무언가가 보이는데 

TfLiteQuantizationParams.scale 와 TfLiteQuantizationParams.zero_point을 이용하면 자동화 가능할 느낌.

template <class T>
void get_top_n(T* prediction, int prediction_size, size_t num_results,
               float threshold, std::vector<std::pair<float, int>>* top_results,
               TfLiteType input_type) {
  // Will contain top N results in ascending order.
  std::priority_queue<std::pair<float, int>, std::vector<std::pair<float, int>>,
                      std::greater<std::pair<float, int>>>
      top_result_pq;

  const long count = prediction_size;  // NOLINT(runtime/int)
  float value = 0.0;

  for (int i = 0; i < count; ++i) {
    switch (input_type) {
      case kTfLiteFloat32:
        value = prediction[i];
        break;
      case kTfLiteInt8:
        value = (prediction[i] + 128) / 256.0;
        break;
      case kTfLiteUInt8:
        value = prediction[i] / 255.0;
        break;
      default:
        break;
    }
    // Only add it if it beats the threshold and has a chance at being in
    // the top N.
    if (value < threshold) {
      continue;
    }

    top_result_pq.push(std::pair<float, int>(value, i));

    // If at capacity, kick the smallest value out.
    if (top_result_pq.size() > num_results) {
      top_result_pq.pop();
    }
  }

  // Copy to output vector and reverse into descending order.
  while (!top_result_pq.empty()) {
    top_results->push_back(top_result_pq.top());
    top_result_pq.pop();
  }
  std::reverse(top_results->begin(), top_results->end());
}

netron 에서 보면 quantization 범위가 나오는데, 어딘가 저장하고는 있는 듯 해서 검색 중

// SupportedQuantizationTypes.
typedef enum TfLiteQuantizationType {
  // No quantization.
  kTfLiteNoQuantization = 0,
  // Affine quantization (with support for per-channel quantization).
  // Corresponds to TfLiteAffineQuantization.
  kTfLiteAffineQuantization = 1,
} TfLiteQuantizationType;

// Structure specifying the quantization used by the tensor, if-any.
typedef struct TfLiteQuantization {
  // The type of quantization held by params.
  TfLiteQuantizationType type;
  // Holds a reference to one of the quantization param structures specified
  // below.
  void* params;
} TfLiteQuantization;

// Legacy. Will be deprecated in favor of TfLiteAffineQuantization.
// If per-layer quantization is specified this field will still be populated in
// addition to TfLiteAffineQuantization.
// Parameters for asymmetric quantization. Quantized values can be converted
// back to float using:
//     real_value = scale * (quantized_value - zero_point)
typedef struct TfLiteQuantizationParams {
  float scale;
  int32_t zero_point;
} TfLiteQuantizationParams;

// Parameters for asymmetric quantization across a dimension (i.e per output
// channel quantization).
// quantized_dimension specifies which dimension the scales and zero_points
// correspond to.
// For a particular value in quantized_dimension, quantized values can be
// converted back to float using:
//     real_value = scale * (quantized_value - zero_point)
typedef struct TfLiteAffineQuantization {
  TfLiteFloatArray* scale;
  TfLiteIntArray* zero_point;
  int32_t quantized_dimension;
} TfLiteAffineQuantization;

typedef struct TfLiteTensor {
  TfLiteType type;
  TfLitePtrUnion data;
  TfLiteIntArray* dims;
  TfLiteQuantizationParams params;
  TfLiteAllocationType allocation_type;
  size_t bytes;
  const void* allocation;
  const char* name;
  struct TfLiteDelegate* delegate;
  TfLiteBufferHandle buffer_handle;
  bool data_is_stale;
  bool is_variable;
  TfLiteQuantization quantization;
  TfLiteSparsity* sparsity;
  const TfLiteIntArray* dims_signature;
} TfLiteTensor;

+

2021.05.26

typedef struct TfLiteIntArray {
  int size;
// gcc 6.1+ have a bug where flexible members aren't properly handled
// https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c
#if (!defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
     __GNUC_MINOR__ >= 1) ||                                      \
    defined(HEXAGON) || (__clang_major__ == 7 && __clang_minor__ == 1)
  int data[0];
#else
  int data[];
#endif
} TfLiteIntArray;

// Fixed size list of floats. Used for per-channel quantization.
typedef struct TfLiteFloatArray {
  int size;
// gcc 6.1+ have a bug where flexible members aren't properly handled
// https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c
// This also applies to the toolchain used for Qualcomm Hexagon DSPs.
#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
    __GNUC_MINOR__ >= 1
  float data[0];
#else
  float data[];
#endif
} TfLiteFloatArray;

class Interpreter {
  /// Invoke the interpreter (run the whole graph in dependency order).
  ///
  /// NOTE: It is possible that the interpreter is not in a ready state
  /// to evaluate (i.e. if a ResizeTensor() has been performed without an
  /// AllocateTensors().
  /// Returns status of success or failure.
  TfLiteStatus Invoke();

  /// WARNING: Experimental interface, subject to change
  Subgraph& primary_subgraph() {
    return *subgraphs_.front();  /// Safe as subgraphs_ always has 1 entry.
  }
  
  /// Read only access to list of inputs.
  const std::vector<int>& inputs() const { return primary_subgraph().inputs(); }
  
  /// Read only access to list of outputs.
  const std::vector<int>& outputs() const {
    return primary_subgraph().outputs();
  }
  
  // Subgraphs
  std::vector<std::unique_ptr<Subgraph>> subgraphs_;
};

[링크 : https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/interpreter.h]

class Subgraph {
  // Array of indices representing the tensors that are inputs to the
  // interpreter.
  std::vector<int> inputs_;

  // Array of indices representing the tensors that are outputs to the
  // interpreter.
  std::vector<int> outputs_;
  
  // Read only access to list of inputs.
  std::vector<int>& inputs() { return inputs_; }

  // Read only access to list of inputs.
  const std::vector<int>& inputs() const { return inputs_; }

  // Read only access to list of outputs.
  std::vector<int>& outputs() { return outputs_; }

  // Read only access to list of outputs.
  const std::vector<int>& outputs() const { return outputs_; }
};

[링크 : https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/core/subgraph.h]

라즈베리 파이 3b+ 에 lxde 올리지 않고 콘솔로만 해서 메모리 최대한 확보하고 쓰는 중인데

비디오 메모리 64MB 정도 먹혔는지 가용 메모리가 706MiB 정도.

그런 이유로 텐서플로우 빌드시 rpi 메모리가 2G 넘지 않으면 코어 하나만 쓰게 되어있다.

그래서 distcc 통해 개별 노드에서 2개씩, 총 3개 노드/ 6 core 쓰도록 하니

35~40분 정도 걸리던 빌드가 8분 40초 만에 끝내준다.

개별 노드에서 병렬시 2개 까지만 허락하도록 해놔야지 안그러면

메모리 부족으로 스왑한다고 io 폭주해서 라즈베리가 먹통이 된다.

/etc/default/distcc

이번 빌드 옵션들을 뒤져보다 보니 빼먹고 안 넣은게 있었네

(라고 쓰고 실은 귀찮았던...)

[링크 : http://stackoverflow.com/questions/14962447/gcc-options-for-a-freescale-imx6q-arm-processor]

[링크 : http://gcc.gnu.org/wiki/FloatingPointMath]

[링크 : http://www.programmersought.com/article/5320739655/]

[링크 : http://www.google.com/amp/s/learnopencv.com/build-and-install-opencv-4-for-raspberry-pi/amp/]

변경된 파일 목록과 함께 보는 방법

[링크 : https://stackoverflow.com/questions/1230084/how-to-have-git-log-show-filenames-like-svn-log-v/1230094]

이번에 한번 날릴 각오로(!) 써봐야지 -_ㅠ

git stash [push]

git stash list

git stash apply | pop

[링크 : https://gmlwjd9405.github.io/2018/05/18/git-stash.html]

%p로 출력하니 0일 경우 (nil)로 출력되네?

[링크 : https://stackoverflow.com/questions/4744650/nil-pointer-in-c-c]

변수 추적해보니 그게 그거인가?

  int output = interpreter->outputs()[0];
  TfLiteIntArray* output_dims = interpreter->tensor(output)->dims;
  // assume output dims to be something like (1, 1, ... ,size)
  auto output_size = output_dims->data[output_dims->size - 1];

    const float* detection_locations = interpreter->tensor(interpreter->outputs()[0])->data.f;
    const float* detection_classes=interpreter->tensor(interpreter->outputs()[1])->data.f;
    const float* detection_scores = interpreter->tensor(interpreter->outputs()[2])->data.f;
    const int    num_detections = *interpreter->tensor(interpreter->outputs()[3])->data.f;

    //there are ALWAYS 10 detections no matter how many objects are detectable
    //cout << "number of detections: " << num_detections << "\n";

    const float confidence_threshold = 0.5;
    for(int i = 0; i < num_detections; i++){
        if(detection_scores[i] > confidence_threshold){
            int  det_index = (int)detection_classes[i]+1;
            float y1=detection_locations[4*i  ]*cam_height;
            float x1=detection_locations[4*i+1]*cam_width;
            float y2=detection_locations[4*i+2]*cam_height;
            float x2=detection_locations[4*i+3]*cam_width;

            Rect rec((int)x1, (int)y1, (int)(x2 - x1), (int)(y2 - y1));
            rectangle(src,rec, Scalar(0, 0, 255), 1, 8, 0);
            putText(src, format("%s", Labels[det_index].c_str()), Point(x1, y1-5) ,FONT_HERSHEY_SIMPLEX,0.5, Scalar(0, 0, 255), 1, 8, 0);
        }
    }

typedef struct {
  int size;
#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
    __GNUC_MINOR__ >= 1
  int data[0];
#else
  int data[];
#endif
} TfLiteIntArray;

typedef union {
  int* i32;
  int64_t* i64;
  float* f;
  char* raw;
  const char* raw_const;
  uint8_t* uint8;
  bool* b;
  int16_t* i16;
  TfLiteComplex64* c64;
  int8_t* int8;
} TfLitePtrUnion;

typedef struct {
  TfLiteType type;
  TfLitePtrUnion data;
  TfLiteIntArray* dims;
  TfLiteQuantizationParams params;
  TfLiteAllocationType allocation_type;
  size_t bytes;
  const void* allocation;
  const char* name;
  TfLiteDelegate* delegate;
  TfLiteBufferHandle buffer_handle;
  bool data_is_stale;
  bool is_variable;
  TfLiteQuantization quantization;
} TfLiteTensor;

[링크 : https://android.googlesource.com/platform/external/tensorflow/.../tensorflow/lite/c/c_api_internal.h]

현재 소스에서는 common.h 로 옮겨진듯

[링크 : https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/c/common.h]

그런 이유로 distcc-pump 모드 시도 ㅠㅠ

[링크 : https://itmir.tistory.com/454]

[링크 : https://www.whatwant.com/entry/Ubuntu에서-ccache-사용하기]