• Hello TensorFlow
• TensorFlow C library
• TensorFlow Go bingding

Hello TensorFlow

人類喜歡將所有事物都納入鄙視鏈的范疇，寵物當(dāng)然也不例外。一般來說，擁有一只純種寵物可以讓主人占據(jù)鄙視鏈的云端，進而鄙視那些混血或者流浪寵物。甚至還發(fā)展出了專業(yè)的鑒定機構(gòu)，可以頒發(fā)《血統(tǒng)證明書》。但是考究各類純種鑒定的常規(guī)方法：例如眼睛的大小、顏色、鼻子的特點、身軀長度、尾巴特征、毛發(fā)等，當(dāng)然也包括一些比較玄幻的特征：寵物家族的個性、氣質(zhì)等等。拋開“黑魔法”不在此討論之外，既然是基于生物外形特征鑒定，判斷是否純種的需求本質(zhì)上就是一個圖像識別服務(wù)。

Tensorflow is not a Machine Learning specific library, instead, is a general purpose computation library that represents computations with graphs.

TensorFlow 開源軟件庫（Apache 2.0 許可證），最初由 Google Brain 團隊開發(fā)。TensorFlow 提供了一系列算法模型和編程接口，讓我們可以快速構(gòu)建一個基于機器學(xué)習(xí)的智能服務(wù)。對于開發(fā)者來說，目前有四種編程接口可供選擇：

• C++ source code: Tensorflow 核心基于 C++ 編寫，支持從高到低各個層級的操作;
• Python bindings & Python library: 對標(biāo) C++ 實現(xiàn)，支持 Python 調(diào)用 C++ 函數(shù);
• Java bindings;
• Go binding;

下面是一個簡單的實例：

[圖片上傳失敗...(image-8754c1-1518459129749)]

環(huán)境準(zhǔn)備

• 安裝 TensorFlow C library,包含一個頭文件 c_api.h 和 libtensorflow.so

wget https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-linux-x86_64-1.5.0.tar.gz

## options
TF_TYPE="cpu" # Change to "gpu" for GPU support
TF_VERSION='1.5.0'
curl -L \
  "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-${TF_TYPE}-$(go env GOOS)-x86_64-${TF_VERSION}.tar.gz" |

• 安裝 Go 語言環(huán)境,參考：玩轉(zhuǎn)編程語言：Golang

• 安裝 Tensorflow Go binding library

go get github.com/tensorflow/tensorflow/tensorflow/go
go get github.com/tensorflow/tensorflow/tensorflow/go/op

• 下載模型（demo model)，包含一個標(biāo)簽文件 label_strings.txt 和 graph.pb

mkdir model
wget https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip -O model/inception.zip
unzip model/inception.zip -d model
chmod -R 777 model

Tensorflow Model Function

//Loading TensorFlow model
func loadModel() error {
  // Load inception model
  model, err := ioutil.ReadFile("./model/tensorflow_inception_graph.pb")
  if err != nil {
    return err
  }
  graph = tf.NewGraph()
  if err := graph.Import(model, ""); err != nil {
    return err
  }
  // Load labels
  labelsFile, err := os.Open("./model/imagenet_comp_graph_label_strings.txt")
  if err != nil {
    return err
  }
  defer labelsFile.Close()
  scanner := bufio.NewScanner(labelsFile)
  // Labels are separated by newlines
  for scanner.Scan() {
    labels = append(labels, scanner.Text())
  }
  if err := scanner.Err(); err != nil {
    return err
  }
  return nil
}

Classifying Workflow

基于 Tensorflow 模型實現(xiàn)圖像識別的主要流程如下：

• 圖像轉(zhuǎn)換 (Convert to tensor )
• 圖像標(biāo)準(zhǔn)化( Normalize )
• 圖像分類 （ Classifying )

func recognizeHandler(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
  // Read image
  imageFile, header, err := r.FormFile("image")
  // Will contain filename and extension
  imageName := strings.Split(header.Filename, ".")
  if err != nil {
    responseError(w, "Could not read image", http.StatusBadRequest)
    return
  }
  defer imageFile.Close()
  var imageBuffer bytes.Buffer
  // Copy image data to a buffer
  io.Copy(&imageBuffer, imageFile)

  // ...

  tensor, err := makeTensorFromImage(&imageBuffer, imageName[:1][0])
  if err != nil {
    responseError(w, "Invalid image", http.StatusBadRequest)
    return
  }

  // ...
}

函數(shù) makeTensorFromImage() which runs an image tensor through the normalization graph.

func makeTensorFromImage(imageBuffer *bytes.Buffer, imageFormat string) (*tf.Tensor, error) {
  tensor, err := tf.NewTensor(imageBuffer.String())
  if err != nil {
    return nil, err
  }
  graph, input, output, err := makeTransformImageGraph(imageFormat)
  if err != nil {
    return nil, err
  }
  session, err := tf.NewSession(graph, nil)
  if err != nil {
    return nil, err
  }
  defer session.Close()
  normalized, err := session.Run(
    map[tf.Output]*tf.Tensor{input: tensor},
    []tf.Output{output},
    nil)
  if err != nil {
    return nil, err
  }
  return normalized[0], nil
}

函數(shù) maketransformimagegraph() 將圖形的像素值調(diào)整到 224x224，以符合模型輸入?yún)?shù)要求。

func makeTransformImageGraph(imageFormat string) (graph *tf.Graph, input, output tf.Output, err error) {
  const (
    H, W  = 224, 224
    Mean  = float32(117)
    Scale = float32(1)
  )
  s := op.NewScope()
  input = op.Placeholder(s, tf.String)
  // Decode PNG or JPEG
  var decode tf.Output
  if imageFormat == "png" {
    decode = op.DecodePng(s, input, op.DecodePngChannels(3))
  } else {
    decode = op.DecodeJpeg(s, input, op.DecodeJpegChannels(3))
  }
  // Div and Sub perform (value-Mean)/Scale for each pixel
  output = op.Div(s,
    op.Sub(s,
      // Resize to 224x224 with bilinear interpolation
      op.ResizeBilinear(s,
        // Create a batch containing a single image
        op.ExpandDims(s,
          // Use decoded pixel values
          op.Cast(s, decode, tf.Float),
          op.Const(s.SubScope("make_batch"), int32(0))),
        op.Const(s.SubScope("size"), []int32{H, W})),
      op.Const(s.SubScope("mean"), Mean)),
    op.Const(s.SubScope("scale"), Scale))
  graph, err = s.Finalize()
  return graph, input, output, err
}

最后，將格式化的 image tensor 輸入到 Inception model graph 中運算。

session, err := tf.NewSession(graph, nil)
if err != nil {
  log.Fatal(err)
}
defer session.Close()
output, err := session.Run(
  map[tf.Output]*tf.Tensor{
    graph.Operation("input").Output(0): tensor,
  },
  []tf.Output{
    graph.Operation("output").Output(0),
  },
  nil)
if err != nil {
  responseError(w, "Could not run inference", http.StatusInternalServerError)
  return
}

Testing

func main() {
  if err := loadModel(); err != nil {
    log.Fatal(err)
    return
  }
  r := httprouter.New()
  r.POST("/recognize", recognizeHandler)
  err := http.ListenAndServe(":8080", r)
  if err != nil {
    log.Println(err)
    return
  }
}

識別案例：黑天鵝

$ curl localhost:8080/recognize -F 'image=@../data/IMG_3560.png'
{
  "filename":"IMG_3000.png",
  "labels":[
    {"label":"black swan","probability":0.98746836,"Percent":"98.75%"},
    {"label":"oystercatcher","probability":0.0040768473,"Percent":"0.41%"},
    {"label":"American coot","probability":0.002185003,"Percent":"0.22%"},
    {"label":"black stork","probability":0.0011524856,"Percent":"0.12%"},
    {"label":"redshank","probability":0.0010183558,"Percent":"0.10%"}]
}

IMG_3560.png

IMG_3608.png

通過上面的案例我們可以發(fā)現(xiàn)，這個服務(wù)目前可以對于黑天鵝圖像的推算概率值為 98.75%，非常準(zhǔn)確；但是對于另外兩張寵物狗的圖像，最高的推算概率值也僅有 30% 左右，雖然也沒有被識別成貓咪或者狼，但是和理想效果要求可用性還有一段距離（此處暫時忽略物種本身的復(fù)雜性）。主要是因為現(xiàn)在我們使用的還只是一個非常“原始”的模型，如果需要為小眾領(lǐng)域服務(wù)（寵物，也可以是其它事物），需要通過訓(xùn)練（Training Models）增強優(yōu)化，或者引入更豐富的標(biāo)簽，更合適的模型。當(dāng)然，訓(xùn)練過程中也會存在樣本質(zhì)量不佳的情況，錯誤樣本和各種噪音也會影響準(zhǔn)確度。

擴展閱讀

• How to Retrain Inception's Final Layer for New Categories

• Getting Started with TensorFlow: A Machine Learning Tutorial

• YoutubeIntroduction To TensorFlow

• We Need to Go Deeper: A Practical Guide to Tensorflow and Inception

• Tensorflow.org: Image Recognition

We know that label 866 (military uniform) should be the top label for the Admiral Hopper image.

擴展閱讀:《The Machine Learning Master》

image

• Machine Learning(一):基于 TensorFlow 實現(xiàn)寵物血統(tǒng)智能識別
• Machine Learning(二):寵物智能識別之 Using OpenCV with Node.js
• Machine Learning:機器學(xué)習(xí)項目
• Machine Learning:機器學(xué)習(xí)算法
• Machine Learning:機器學(xué)習(xí)書單
• Machine Learning:人工智能媒體報道集
• Machine Learning:機器學(xué)習(xí)技術(shù)與知識產(chǎn)權(quán)法
• Machine Learning:經(jīng)濟學(xué)家談人工智能
• 數(shù)據(jù)可視化（三）基于 Graphviz 實現(xiàn)程序化繪圖