TensorFlow2.8.0代码分析之例子wav_to_spectrogram\wav_to_spectrogram之WavToSpectrogram函数

// Runs a TensorFlow graph to convert an audio file into a visualization.
tensorflow::Status WavToSpectrogram(const tensorflow::string& input_wav,
                                    int32_t window_size, int32_t stride,
                                    float brightness,
                                    const tensorflow::string& output_image)

函数作用：使用 TensorFlow graph，将一音频文件可视化。

 函数调用函数情况如下：

 

函数声明情况如下：

 

函数声明文件如下：

 

函数流程图如下：

函数逻辑顺序图如下：

函数原始代码如下：

// Runs a TensorFlow graph to convert an audio file into a visualization.图形化一个音频

tensorflow::Status WavToSpectrogram(const tensorflow::string& input_wav,

                                    int32_t window_size, int32_t stride,

                                    float brightness,

                                    const tensorflow::string& output_image) {

  auto root = tensorflow::Scope::NewRootScope();

  using namespace tensorflow::ops;  // NOLINT(build/namespaces)

  // The following block creates a TensorFlow graph that:创建Tensoflow graph

  //  - Reads and decodes the audio file into a tensor of float samples.读入并解码音频到float sample

  //  - Creates a float spectrogram from those samples.从样品中创建float spectrogram

  //  - Scales, clamps, and converts that spectrogram to 0 to 255 uint8's.转换归一化

  //  - Reshapes the tensor so that it's [height, width, 1] for imaging.整形

  //  - Encodes it as a PNG stream and saves it out to a file.PNG编码

  Output file_reader =

      tensorflow::ops::ReadFile(root.WithOpName("input_wav"), input_wav);

  DecodeWav wav_decoder =

      DecodeWav(root.WithOpName("wav_decoder"), file_reader);

  Output spectrogram = AudioSpectrogram(root.WithOpName("spectrogram"),

                                        wav_decoder.audio, window_size, stride);

  Output brightness_placeholder =

      Placeholder(root.WithOpName("brightness_placeholder"), DT_FLOAT,

                  Placeholder::Attrs().Shape(TensorShape({})));

  Output mul = Mul(root.WithOpName("mul"), spectrogram, brightness_placeholder);

  Output min_const = Const(root.WithOpName("min_const"), 255.0f);

  Output min = Minimum(root.WithOpName("min"), mul, min_const);

  Output cast = Cast(root.WithOpName("cast"), min, DT_UINT8);

  Output expand_dims_const = Const(root.WithOpName("expand_dims_const"), -1);

  Output expand_dims =

      ExpandDims(root.WithOpName("expand_dims"), cast, expand_dims_const);

  Output squeeze = Squeeze(root.WithOpName("squeeze"), expand_dims,

                           Squeeze::Attrs().Axis({0}));

  Output png_encoder = EncodePng(root.WithOpName("png_encoder"), squeeze);

  tensorflow::ops::WriteFile file_writer = tensorflow::ops::WriteFile(

      root.WithOpName("output_image"), output_image, png_encoder);

  tensorflow::GraphDef graph;

  TF_RETURN_IF_ERROR(root.ToGraphDef(&graph));

  // Build a session object from this graph definition.从图形定义建立对话 The power of TensorFlow

  // is that you can reuse complex computations like this我们可以重用复杂的计算, so usually we'd run a

  // lot of different inputs through it. In this example, we're just doing a

  // one-off run, so we'll create it and then use it immediately.

  std::unique_ptr<tensorflow::Session> session(

      tensorflow::NewSession(tensorflow::SessionOptions()));

  TF_RETURN_IF_ERROR(session->Create(graph));

  // We're passing in the brightness as an input, so create a tensor to hold the

  // value.使用亮度作为输入

  tensorflow::Tensor brightness_tensor(DT_FLOAT, TensorShape({}));

  brightness_tensor.scalar<float>()() = brightness;

  // Run the session to analyze the audio and write out the file.分析音频并输出文件

  TF_RETURN_IF_ERROR(

      session->Run({{"brightness_placeholder", brightness_tensor}}, {},

                   {"output_image"}, nullptr));

  return tensorflow::Status::OK();

}