記錄Tensorflow 搭建過程，基於Docker

MNIST 模型計算

Tensorflow Serveing 搭建過程，基於Docker

模型測試

Tensorflow-Serveing gRPC 遠端預測呼叫

Tensorflow -Serveing RESTful 遠端預測呼叫

Tensorflow 基於Docker 搭建過程

版本使用說明：

docker 19.03.4

Tensorflow:1.12.3-py3 使用映象：tensorflow/tensorflow:1.12.3-py3

映象包括：

- Jupyter Notebook- TensorFlow- scikit-learn- pandas- matplotlib- numpy- scipy- Pillow- Python3

安裝Docker環境啟動 啟動tensorflowdocker run --name notebooks -d -v /home/notebooks:/notebooks -v /home/logs:/logs -p 8888:8888 tensorflow/tensorflow:1.12.3-py3 /run_jupyter.sh --allow-root --NotebookApp.token='abc123'啟動TensorBoarddocker run --name board -d -v /home/logs:/logs -p 6006:6006 tensorflow/tensorflow:1.12.3-py3 tensorboard --logdir /logs

      開啟：http://<ip>:8888  和 http://<ip>:6006 開啟jupyter 和 board 

MNIST模型訓練並生成model上傳mnist 對應的程式碼。避免下載訓練資料失敗， 這裡一併上傳。進入notebook的容器， 執行訓練資料

python /notebooks/mnist/source/mnist_saved_model.py --training_iteration=5000 --model_version=1 --work_dir=/notebooks/mnist/data /notebooks/model

3. 訓練完後 ， 資料被儲存在了/notebooks/model中對應宿主機的/home/notebooks/model中

啟動tensorflow-serveing docker 並指定model位置

docker run --rm -p 8501:8501 -p 8500:8500 \-v "/home/notebooks/model:/models/mnist" \-e MODEL_NAME=mnist -t tensorflow/serving &

說明：8500 gRPC使用埠， 8501 RESTful使用埠

mnist 模型名稱

模型測試

啟動後，使用mnist_client估算模型(因為需要tensorflow環境，這裡直接進入notebook的容器執行)

python ./mnist/source/mnist_client.py --server=172.17.0.3:8500 --work_dir=./mnist/data

執行中會提示缺少tensorflow_serving的模組。

因為我們使用的是1.12.3的tensorflow的版本，這邊需要使用pip下載該模組

pip install 'tensorflow-serving-api~=1.12.3'

到此tensorflow-serveing 已經啟動，相關內容可以參考：https://www.tensorflow.org/tfx/serving/api_rest

http://<ip>:8501/v1/models/mnist/metadata 檢視metadata定義

外部訪問tensorflow-serveing gRPC方式識別

參考“TensorFlow 建立過程（單機版）” 可以建立單機版本的Tensorflow+notebook 的環境， 並透過該環境執行py檔案對圖片進行識別。 conda 啟動tensorflow12 的環境， 啟動jupyter

以下為呼叫的方法

import timeimport numpy as npimport grpcfrom tensorflow.contrib.util import make_tensor_protofrom tensorflow.contrib import util as contrib_utilfrom tensorflow_serving.apis import predict_pb2from tensorflow_serving.apis import prediction_service_pb2_grpcimport imageioimport matplotlib.pyplot as pltfrom PIL import Imagedef run(host, port, image, model, signature_name):    channel = grpc.insecure_channel('{host}:{port}'.format(host=host, port=port))    stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)        # Read an image    data = imageio.imread(image)    data = data.astype(np.float32)    #data = imageprepare(image)        data1 = Image.open(image);    plt.imshow(data1)    plt.axis('off') # 不顯示座標軸    plt.show()    #print(data)        start = time.time()        # Call classification model to make prediction on the image    request = predict_pb2.PredictRequest()    request.model_spec.name = model    request.model_spec.signature_name = signature_name    #mtp = make_tensor_proto(data, shape=[1, 784])    mtp = contrib_util.make_tensor_proto(data, shape=[1, 784])    request.inputs['images'].CopyFrom(mtp)    #print(request)    result = stub.Predict(request, 10.0)        end = time.time()    time_diff = end - start        print('耗時: {}'.format(time_diff))    response = np.array(result.outputs['scores'].float_val)    prediction = np.argmax(response)    print('{}{}'.format('識別為:',prediction))

run("yun.mydomain.top", 8500, "./data/1.png", "mnist", "predict_images")

run("yun.mydomain.top", 8500, "./data/2.png", "mnist", "predict_images")

run("yun.mydomain.top", 8500, "./data/3.png", "mnist", "predict_images")

run("yun.mydomain.top", 8500, "./data/4.png", "mnist", "predict_images")

run("yun.mydomain.top", 8500, "./data/5.png", "mnist", "predict_images")

run("yun.mydomain.top", 8500, "./data/6.png", "mnist", "predict_images")

run("yun.mydomain.top", 8500, "./data/7.png", "mnist", "predict_images")

run("yun.mydomain.top", 8500, "./data/8.png", "mnist", "predict_images")

run("yun.mydomain.top", 8500, "./data/9.png", "mnist", "predict_images")

外部透過RESTful API的方式呼叫

以下透過python的方法來使用，因為各種語言均可使用RESTful的呼叫。 則Python透過其他也可以一樣透過。

import matplotlib.pyplot as pltimport imageioimport numpy as npfrom PIL import Imageimport requestsimport jsonimport timedef run(host, port, image_file, model, signature_name):    image = Image.open(image_file)    #resized_image = image.resize((28, 28), Image.ANTIALIAS)        data = np.array(Image.open(image_file).convert('L').resize((28, 28))).astype(np.float).reshape(-1, 28, 28, 1)        #data =resized_image# imageio.imread(resized_image)    plt.imshow(image)    plt.axis('off') # 不顯示座標軸    plt.show()    data = data.astype(np.float32)    data = data.flatten() #變成1維陣列    t= np.array2string(data, separator=',', formatter={'float':lambda x: "%f" % x})    json_request = '{{"signature_name": "{}","inputs":{{"images":[{}] }}}}'.format(signature_name,t)    start = time.time()    resp = requests.post('http://{}:{}/v1/models/{}:predict'.format(host,port,model), data=json_request)    res = json.loads(resp.content)    end = time.time()    time_diff = end - start    print('耗時: {}'.format(time_diff))    response = np.array(res['outputs'])    prediction = np.argmax(response)    print('{}{}'.format('識別為:',prediction))

run("yun.mydomain.top", 8501, "./data/1.png", "mnist", "predict_images")

run("yun.mydomain.top", 8501, "./data/2.png", "mnist", "predict_images")

run("yun.mydomain.top", 8501, "./data/3.png", "mnist", "predict_images")

run("yun.mydomain.top", 8501, "./data/4.png", "mnist", "predict_images")

run("yun.mydomain.top", 8501, "./data/5.png", "mnist", "predict_images")

run("yun.mydomain.top", 8501, "./data/6.png", "mnist", "predict_images")

run("yun.mydomain.top", 8501, "./data/7.png", "mnist", "predict_images")

run("yun.mydomain.top", 8501, "./data/8.png", "mnist", "predict_images")

run("yun.zhucl1006.top", 8501, "./data/9.png", "mnist", "predict_images")