//一个最小的但有用的C++的例子展示了如何加载一个ImageNet式目标识别tensorflow模型,准备输入图像,通过图运行和解释结果。
//它的目的是有尽可能少的依赖和尽可能清楚,所以它比它在产品代码中更详细。特别是,自动使用
//TensorFlow的大量返回值的类型能删除大量的样板,但我发现在样本代码中,显式类型是有用的,使其查找相关的类变得简单。
//要使用它,在工作目录中learning/brain/tutorials/label_image /data/文件夹下面编译并运行,你应该能看到Lena图像这个例子输出的前五个标签。然后你可以自定义它使用您自己的模型或图像,通过在main()函数中改变文件名。
//默认包括的googlenet_graph.pb文件是创建自Inception。
#include <fstream>
#include <vector>
#include "tensorflow/cc/ops/const_op.h"
#include "tensorflow/cc/ops/image_ops.h"
#include "tensorflow/cc/ops/standard_ops.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/graph/default_device.h"
#include "tensorflow/core/graph/graph_def_builder.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/stringpiece.h"
#include "tensorflow/core/lib/core/threadpool.h"
#include "tensorflow/core/lib/io/path.h"
#include "tensorflow/core/lib/strings/stringprintf.h"
#include "tensorflow/core/platform/init_main.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/public/session.h"
#include "tensorflow/core/util/command_line_flags.h"
// 这些都是公共类,它很方便的引用没有命名空间。
using tensorflow::Flag;
using tensorflow::Tensor;
using tensorflow::Status;
using tensorflow::string;
using tensorflow::int32;
// 取一个文件名,从它加载一个标签列表,每行一个,返回一个字符串的向量。它填充空字符串,所以结果的长度是16的倍数,因为我们的模型期望这样。
Status ReadLabelsFile(string file_name, std::vector<string>* result,
size_t* found_label_count) {
std::ifstream file(file_name);
if (!file) {
return tensorflow::errors::NotFound("Labels file ", file_name,
" not found.");
}
result->clear();
string line;
while (std::getline(file, line)) {
result->push_back(line);
}
*found_label_count = result->size();
const int padding = 16;
while (result->size() % padding) {
result->emplace_back();
}
return Status::OK();
}
// 给定一个图像文件名,读入数据,尝试将其解码为图像,将它调整到所要求的大小,然后按需要缩放值。
Status ReadTensorFromImageFile(string file_name, const int input_height,
const int input_width, const float input_mean,
const float input_std,
std::vector<Tensor>* out_tensors) {
auto root = tensorflow::Scope::NewRootScope();
using namespace ::tensorflow::ops; // NOLINT(build/namespaces)
string input_name = "file_reader";
string output_name = "normalized";
auto file_reader = tensorflow::ops::ReadFile(root.WithOpName(input_name),
file_name);
// 现在尝试找出它是什么样的文件,并解码它。
const int wanted_channels = 3;
tensorflow::Output image_reader;
if (tensorflow::StringPiece(file_name).ends_with(".png")) {
image_reader = DecodePng(root.WithOpName("png_reader"), file_reader,
DecodePng::Channels(wanted_channels));
} else if (tensorflow::StringPiece(file_name).ends_with(".gif")) {
image_reader = DecodeGif(root.WithOpName("gif_reader"), file_reader);
} else {
// 假设图像不是 PNG 或是 GIF 格式,就一定是JPEG格式.
image_reader = DecodeJpeg(root.WithOpName("jpeg_reader"), file_reader,
DecodeJpeg::Channels(wanted_channels));
}
// 现在把图像数据cast为float类型,这样我们可以做常规的数学计算.
auto float_caster =
Cast(root.WithOpName("float_caster"), image_reader, tensorflow::DT_FLOAT);
// 在TensorFlow里,图像ops的约定是所有图像都能批处理,以使它们是具有[批,高度,宽度,通道]索引的四维数组。因为我们只有一个图像,我们
//用expanddims()函数开始添加1维的批尺寸。
auto dims_expander = ExpandDims(root, float_caster, 0);
// 双线调整图像大小以适合其要求的尺寸。
auto resized = ResizeBilinear(
root, dims_expander,
Const(root.WithOpName("size"), {input_height, input_width}));
// 减去mean,除以scale
Div(root.WithOpName(output_name), Sub(root, resized, {input_mean}),
{input_std});
// 这个运行我们刚刚建的GraphDef网络模型的定义,返回输出张量的结果
tensorflow::GraphDef graph;
TF_RETURN_IF_ERROR(root.ToGraphDef(&graph));
std::unique_ptr<tensorflow::Session> session(
tensorflow::NewSession(tensorflow::SessionOptions()));
TF_RETURN_IF_ERROR(session->Create(graph));
TF_RETURN_IF_ERROR(session->Run({}, {output_name}, {}, out_tensors));
return Status::OK();
}
// 从磁盘读取一个模型图像的定义,创建session对象,你可以使用它来运行
Status LoadGraph(string graph_file_name,
std::unique_ptr<tensorflow::Session>* session) {
tensorflow::GraphDef graph_def;
Status load_graph_status =
ReadBinaryProto(tensorflow::Env::Default(), graph_file_name, &graph_def);
if (!load_graph_status.ok()) {
return tensorflow::errors::NotFound("Failed to load compute graph at '",
graph_file_name, "'");
}
session->reset(tensorflow::NewSession(tensorflow::SessionOptions()));
Status session_create_status = (*session)->Create(graph_def);
if (!session_create_status.ok()) {
return session_create_status;
}
return Status::OK();
}
// 分析Inception图像的输出,以检索最高分数和它们在张量中的位置,这些位置对应于类别。
Status GetTopLabels(const std::vector<Tensor>& outputs, int how_many_labels,
Tensor* indices, Tensor* scores) {
auto root = tensorflow::Scope::NewRootScope();
using namespace ::tensorflow::ops; // NOLINT(build/namespaces)
string output_name = "top_k";
TopK(root.WithOpName(output_name), outputs[0], how_many_labels);
// 这个运行我们刚刚建的GraphDef网络模型的定义,返回输出张量的结果
tensorflow::GraphDef graph;
TF_RETURN_IF_ERROR(root.ToGraphDef(&graph));
std::unique_ptr<tensorflow::Session> session(
tensorflow::NewSession(tensorflow::SessionOptions()));
TF_RETURN_IF_ERROR(session->Create(graph));
// TopK节点返回两个输出,分数和原来的索引,因此,我们必须追加:0和:1指定两者。
std::vector<Tensor> out_tensors;
TF_RETURN_IF_ERROR(session->Run({}, {output_name + ":0", output_name + ":1"},
{}, &out_tensors));
*scores = out_tensors[0];
*indices = out_tensors[1];
return Status::OK();
}
// 给定一个模型运行的输出,以及包含该标签的文件的名称打印出得分最高值的前五名。
Status PrintTopLabels(const std::vector<Tensor>& outputs,
string labels_file_name) {
std::vector<string> labels;
size_t label_count;
Status read_labels_status =
ReadLabelsFile(labels_file_name, &labels, &label_count);
if (!read_labels_status.ok()) {
LOG(ERROR) << read_labels_status;
return read_labels_status;
}
const int how_many_labels = std::min(5, static_cast<int>(label_count));
Tensor indices;
Tensor scores;
TF_RETURN_IF_ERROR(GetTopLabels(outputs, how_many_labels, &indices, &scores));
tensorflow::TTypes<float>::Flat scores_flat = scores.flat<float>();
tensorflow::TTypes<int32>::Flat indices_flat = indices.flat<int32>();
for (int pos = 0; pos < how_many_labels; ++pos) {
const int label_index = indices_flat(pos);
const float score = scores_flat(pos);
LOG(INFO) << labels[label_index] << " (" << label_index << "): " << score;
}
return Status::OK();
}
// 这是一个测试函数,返回最顶上的标签索引是否为预期的。
Status CheckTopLabel(const std::vector<Tensor>& outputs, int expected,
bool* is_expected) {
*is_expected = false;
Tensor indices;
Tensor scores;
const int how_many_labels = 1;
TF_RETURN_IF_ERROR(GetTopLabels(outputs, how_many_labels, &indices, &scores));
tensorflow::TTypes<int32>::Flat indices_flat = indices.flat<int32>();
if (indices_flat(0) != expected) {
LOG(ERROR) << "Expected label #" << expected << " but got #"
<< indices_flat(0);
*is_expected = false;
} else {
*is_expected = true;
}
return Status::OK();
}
int main(int argc, char* argv[]) {
// 他们定义图形和输入数据的位置,以及什么样的输入模型是期望的。
//如果你训练自己的模型,或使用GoogLeNet以外的模型,你需要更新这些。
string image = "tensorflow/examples/label_image/data/grace_hopper.jpg";
string graph =
"tensorflow/examples/label_image/data/"
"tensorflow_inception_graph.pb";
string labels =
"tensorflow/examples/label_image/data/"
"imagenet_comp_graph_label_strings.txt";
int32 input_width = 299;
int32 input_height = 299;
int32 input_mean = 128;
int32 input_std = 128;
string input_layer = "Mul";
string output_layer = "softmax";
bool self_test = false;
string root_dir = "";
std::vector<Flag> flag_list = {
Flag("image", &image, "image to be processed"),
Flag("graph", &graph, "graph to be executed"),
Flag("labels", &labels, "name of file containing labels"),
Flag("input_width", &input_width, "resize image to this width in pixels"),
Flag("input_height", &input_height,
"resize image to this height in pixels"),
Flag("input_mean", &input_mean, "scale pixel values to this mean"),
Flag("input_std", &input_std, "scale pixel values to this std deviation"),
Flag("input_layer", &input_layer, "name of input layer"),
Flag("output_layer", &output_layer, "name of output layer"),
Flag("self_test", &self_test, "run a self test"),
Flag("root_dir", &root_dir,
"interpret image and graph file names relative to this directory"),
};
string usage = tensorflow::Flags::Usage(argv[0], flag_list);
const bool parse_result = tensorflow::Flags::Parse(&argc, argv, flag_list);
if (!parse_result) {
LOG(ERROR) << usage;
return -1;
}
// 我们需要调用这个函数建立Tensorflow的通用状态
tensorflow::port::InitMain(argv[0], &argc, &argv);
if (argc > 1) {
LOG(ERROR) << "Unknown argument " << argv[1] << "\n" << usage;
return -1;
}
// 首先加载和初始化模型
std::unique_ptr<tensorflow::Session> session;
string graph_path = tensorflow::io::JoinPath(root_dir, graph);
Status load_graph_status = LoadGraph(graph_path, &session);
if (!load_graph_status.ok()) {
LOG(ERROR) << load_graph_status;
return -1;
}
// 用float数组的个数从硬盘获取图像,调整大小和归一化到主图像期望的规格
std::vector<Tensor> resized_tensors;
string image_path = tensorflow::io::JoinPath(root_dir, image);
Status read_tensor_status =
ReadTensorFromImageFile(image_path, input_height, input_width, input_mean,
input_std, &resized_tensors);
if (!read_tensor_status.ok()) {
LOG(ERROR) << read_tensor_status;
return -1;
}
const Tensor& resized_tensor = resized_tensors[0];
// 通过model运行图像
std::vector<Tensor> outputs;
Status run_status = session->Run({{input_layer, resized_tensor}},
{output_layer}, {}, &outputs);
if (!run_status.ok()) {
LOG(ERROR) << "Running model failed: " << run_status;
return -1;
}
// 这是用于自动化测试,以确保我们由默认设置得到预期的结果。我们知道标签866(军装)应该是Asmiral Hopper图像的置顶标签。
if (self_test) {
bool expected_matches;
Status check_status = CheckTopLabel(outputs, 866, &expected_matches);
if (!check_status.ok()) {
LOG(ERROR) << "Running check failed: " << check_status;
return -1;
}
if (!expected_matches) {
LOG(ERROR) << "Self-test failed!";
return -1;
}
}
// 对我们产生的结果做一些有意思的事情
Status print_status = PrintTopLabels(outputs, labels);
if (!print_status.ok()) {
LOG(ERROR) << "Running print failed: " << print_status;
return -1;
}
return 0;
}