Question

我已经用python和使用capi编写了最少的代码。此代码加载premade图并通过它运行测试图像。使用capi似乎无法获得与通过python运行相同的结果。 Linux Ubuntu 16.04 我已经安装了tf并使用bazel构建了库。我尝试了各种版本，但为此实验将其固定在1.9。

Python 3.5.2
bazel 0.13.0
gcc 5.4
no cuda

我正在尝试运行android移动示例提供的图表ssd_mobilenet_v1_android_export.pb。

我正在使用this test image

我有简单的python代码（下面的代码），该代码可加载图形，向其提供图像并读取各种张量。此图的末尾有一个称为num_detections的操作，我正在阅读张量以检查是否成功。使用dog.jpg，我得到3.0的结果。其他测试图像会给出适当的结果。

我正在用capi（下面的代码）重复该实验，但是我得到的唯一结果是0。我看的第一个明显的地方是我的图像输入。第二个图形op是ToFloat操作。我一直在用它来检查进入提要的像素数据的值。它们是相同的。解码后，我检查图像没有翻转并且颜色通道的顺序相同。然后，我开始进一步查找该图，在每个操作中采样张量的值。我发现它们几乎是相同的，直到它们进入一长串的卷积运算，似乎值逐渐趋于不同为止-就像它们在前几个运算元中一样，然后开始显示出越来越远的差异。

不可用的验证码：

#include <png.hpp>
#include <stdlib.h>
#include <stdio.h>
#include <tensorflow/c/c_api.h>

typedef unsigned char byte; 

const int num_channels = 3;

struct Image {
  byte* data;
  int width, height;
  int dataSize() {
    return width * height * num_channels;
  }
};

Image open_image(const char* filename) {

  Image image;
  int dataSize = image.width * image.height * num_channels;

  png::image< png::rgb_pixel > png_import(filename);
  image.width = png_import.get_width();
  image.height = png_import.get_height();
  image.data = (byte*)malloc(image.dataSize());
  for(int y=0; y<image.height; y++) {
    for(int x=0; x<image.width; x++) {
      png::rgb_pixel pixel = png_import.get_pixel(x,y);
      byte* data = &image.data[(y*image.width+x)*num_channels];
      *data++ = pixel.red;
      *data++ = pixel.green;
      *data++ = pixel.blue;
    }
  }

  printf("%s loaded, %dx%d\n", filename, image.width, image.height);
  return image;
}

TF_Graph* open_graph(const char* filename, TF_Status* status) {

  TF_Graph* graph = TF_NewGraph();

  TF_ImportGraphDefOptions* opts = TF_NewImportGraphDefOptions();

  FILE* fp = fopen(filename, "r");

  fseek(fp, 0L, SEEK_END);
  int file_size = ftell(fp);
  rewind(fp);

  char* graph_content = (char*)malloc(file_size);

  fread(graph_content, file_size, 1, fp);
  fclose(fp);

  TF_Buffer tfBuffer;
  tfBuffer.data = graph_content;
  tfBuffer.length = file_size;

  TF_GraphImportGraphDef(graph, &tfBuffer, opts, status);

  free(graph_content);

  if(TF_GetCode(status) != TF_OK) {

    printf("%s\n", TF_Message(status));
    return 0;
  }

  printf("Successfully loaded graph\n");

  return graph;
}
void TensorDeallocator(void* data, size_t len, void* arg)
{

}


int main(int argi, char** argc) {

  printf("TensorFlow C library version %s\n", TF_Version());

  TF_Status* status = TF_NewStatus();
  TF_Graph* graph = open_graph(argc[1], status);

  Image image = open_image(argc[2]);

  int64_t dims[] = {1, image.width, image.height, num_channels};

  TF_Tensor* image_tensor = TF_NewTensor(
    TF_UINT8, dims, 4, 
    image.data, image.dataSize(), 
    TensorDeallocator, 0
  );

  TF_Operation* input_image_op = TF_GraphOperationByName(graph, "image_tensor");
  if(!input_image_op) {
    printf("Failed to find Op '%s'", "image_tensor");
    return 1;
  }

  TF_Output image_input;
  image_input.oper = input_image_op;
  image_input.index = 0;
  TF_Output inputs[1] = {image_input};

  TF_Operation* num_detection_op = TF_GraphOperationByName(graph, "num_detections");
  if(!num_detection_op) {
    printf("Failed to find Op '%s'", "num_detections");
    return 1;
  }
  TF_Output num_detection_output;
  num_detection_output.oper = num_detection_op;
  num_detection_output.index = 0;

  TF_Output outputs[1] = {num_detection_output};
  TF_Tensor* output_tensors[1];

  TF_SessionOptions * options = TF_NewSessionOptions();
  TF_Session* session = TF_NewSession( graph, options, status );

  TF_SessionRun(session, 0, 
                &image_input, &image_tensor, 1,
                outputs, output_tensors, 1,
                0, 0,
                0, status
                );

  if(TF_GetCode(status) != TF_OK) {

    printf("%s\n", TF_Message(status));
  }
  else {

    printf("Ran successfully\n");

    float* f = (float*)TF_TensorData(output_tensors[0]);

    printf("TF data %f\n", f[0]);

  }

  free(image.data);

  return 0;
}

有效的python代码：

import data_helpers
import sys
import tensorflow as tf
import numpy as np


from tensorflow.python.platform import gfile
from tensorflow.python.util import compat
from tensorflow.core.protobuf import saved_model_pb2

def read_graph(model_filename):

    with gfile.FastGFile(model_filename, 'rb') as f:

        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
        g_in = tf.import_graph_def(graph_def)

        return g_in


def printOps():

    for op in tf.get_default_graph().get_operations():

        print(op.name)

if __name__ == "__main__":

    if len(sys.argv) < 3:
        print("Usage: run_image_on_graph.py [input graph filepath] [output image filepath]")
        exit(1)

    graph_file = sys.argv[1]
    image_file = sys.argv[2]

    image = data_helpers.import_image(image_file)

    with tf.Session() as sess:

        read_graph(graph_file)

        graph = tf.get_default_graph()

        image_tensor = graph.get_tensor_by_name("import/image_tensor:0")
        num_detections = graph.get_tensor_by_name("import/num_detections:0")
        detection_scores = graph.get_tensor_by_name("import/detection_scores:0")
        detection_boxes = graph.get_tensor_by_name("import/detection_boxes:0")
        debug = graph.get_tensor_by_name("import/ToFloat:0")

        image = np.array([data_helpers.import_image(image_file)])

        print("sess = " + str(sess.run([num_detections], feed_dict={
            image_tensor: image})
        ))

我的方法明显有什么问题吗？

使用capi复制结果麻烦

0 个答案: