我的目标是使用tensorflow serving 用docker部署模型后,将服务暴露出来,分别在Python和Java中对模型进行访问,因为tensorflow serving的文档较少,grpc使用花了不少时间,不过总算是可以用了。
后续优化:这样简单地部署的Serving服务,,所以每次调用都需要花比较多的时间(感觉像是每次都需要加载模型,本地加载完模型后单预测只需要十几毫秒),需要后续找时间看看有没有办法让模型预加载,服务调用时使用预测方法。
我直接用tensorflow serving docker部署的,直接按照官方的文档即可,唯一可能不同的是国内的网络问题,可以将下载和安装的步骤从dockerfile里面转移到登陆docker container去手动做。
我的总体环境:
tensorflow 1.3.0
python 3.5
java 1.8
这里我训练了一个分类器,主要有三个分类,主要代码如下:
#设置导出时的目录特征名
export_time = time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime(time.time()))
#为了接收平铺开的图片数组(Java处理比较麻烦) 150528 = 224*224*3
x = tf.placeholder(tf.float32, [None, 150528])
x2 = tf.reshape(x, [-1, 224, 224, 3])
#我自己的网络预测
prob = net.network(x2)
sess = tf.Session()
#恢复模型参数
saver = tf.train.Saver()
module_file = tf.train.latest_checkpoint(weights_path)
saver.restore(sess, module_file)
#获取top 1预测
values, indices = tf.nn.top_k(prob, 1)
#创建模型输出builder
builder = tf.saved_model.builder.SavedModelBuilder(exporter_path + export_time)
#转化tensor到模型支持的格式tensor_info,下面的reshape是因为只想输出单个结果数组,否则是二维的
tensor_info_x = tf.saved_model.utils.build_tensor_info(x)
tensor_info_pro = tf.saved_model.utils.build_tensor_info(tf.reshape(values, [1]))
tensor_info_classify = tf.saved_model.utils.build_tensor_info(tf.reshape(indices, [1]))
#定义方法名和输入输出
signature_def_map = {
"predict_image": tf.saved_model.signature_def_utils.build_signature_def(
inputs={"image": tensor_info_x},
outputs={
"pro": tensor_info_pro,
"classify": tensor_info_classify
},
method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME
)}
builder.add_meta_graph_and_variables(sess,
[tf.saved_model.tag_constants.SERVING],
signature_def_map=signature_def_map)
builder.save()
可以使用python命令生成模型文件夹,里面包含了saved_model.pb文件和variables文件夹
接着在container中可以新建一个文件夹,如serving-models,在文件夹下新建该模型文件夹classify_data,用来存放的模型文件夹,使用docker拷贝的命令拷贝模型到模型文件夹中:
docker cp 本机模型文件夹 containerId:/serving-models/classify_data/模型版本号
启动模型服务,监听9000端口:
bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server --port=9000 --model_name=classify_data --model_base_path=/serving-models/classify_data/
编写Python访问客户端,可以运行看看之前保存模型时,signature_def_map的输入输出:
inputs {
key: "image"
value {
name: "Placeholder:0"
dtype: DT_FLOAT
tensor_shape {
dim {
size: -1
}
dim {
size: 224
}
dim {
size: 224
}
dim {
size: 3
}
}
}
}
outputs {
key: "classify"
value {
name: "ToFloat_1:0"
dtype: DT_FLOAT
tensor_shape {
dim {
size: -1
}
dim {
size: 1
}
}
}
}
outputs {
key: "pro"
value {
name: "TopKV2:0"
dtype: DT_FLOAT
tensor_shape {
dim {
size: -1
}
dim {
size: 1
}
}
}
}
我们可以定义自己的proto文件,并使用tenserflow/serving/api中的proto来生成代码,这里我不打算如此做,而是用pip install tensorflow-serving-client安装了一个第三方提供的库来访问tensorflow serving服务,python代码如下:
import sys
sys.path.insert(0, "./")
from tensorflow_serving_client.protos import predict_pb2, prediction_service_pb2
import cv2
from grpc.beta import implementations
import tensorflow as tf
from tensorflow.python.framework import dtypes
import time
#注意,如果在windows下测试,文件名可能需要写成:im_name = r"测试文件目录\文件名"
im_name = "测试文件目录/文件名"
if __name__ == '__main__':
#文件读取和处理
im = cv2.imread(im_name)
re_im = cv2.resize(im, (224, 224), interpolation=cv2.INTER_CUBIC)
#记个时
start_time = time.time()
#建立连接
channel = implementations.insecure_channel("你的ip", 9000)
stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
request = predict_pb2.PredictRequest()
#这里由保存和运行时定义,第一个是运行时配置的模型名,第二个是保存时输入的方法名
request.model_spec.name = "classify_data"
#入参参照入参定义
request.inputs["image"].ParseFromString(tf.contrib.util.make_tensor_proto(re_im, dtype=dtypes.float32, shape=[1, 224, 224, 3]).SerializeToString())
#第二个参数是最大等待时间,因为这里是block模式访问的
response = stub.Predict(request, 10.0)
results = {}
for key in response.outputs:
tensor_proto = response.outputs[key]
nd_array = tf.contrib.util.make_ndarray(tensor_proto)
results[key] = nd_array
print("cost %ss to predict: " % (time.time() - start_time))
print(results["pro"])
print(results["classify"])
最终输出,例如:
cost 5.115269899368286s to predict:
[ 1.]
[2]
Java和Python一样,可以选择自己编译proto文件,也可以像我一样用第三方库,我是用的是这个http://mvnrepository.com/artifact/com.yesup.oss/tensorflow-client/1.4-2
在pom.xml下加入依赖:
<dependency>
<groupId>com.yesup.ossgroupId>
<artifactId>tensorflow-clientartifactId>
<version>1.4-2version>
dependency>
<dependency>
<groupId>net.coobirdgroupId>
<artifactId>thumbnailatorartifactId>
<version>0.4.8version>
dependency>
<dependency>
<groupId>io.grpcgroupId>
<artifactId>grpc-nettyartifactId>
<version>1.7.0version>
dependency>
<dependency>
<groupId>io.nettygroupId>
<artifactId>netty-tcnative-boringssl-staticartifactId>
<version>2.0.7.Finalversion>
dependency>
Java代码如下:
String file = "文件地址"
//读取文件,强制修改图片大小,设置输出文件格式bmp(模型定义时输入数据是无编码的)
BufferedImage im = Thumbnails.of(file).forceSize(224, 224).outputFormat("bmp").asBufferedImage();
//转换图片到图片数组,匹配输入数据类型为Float
Raster raster = im.getData();
List floatList = new ArrayList<>();
float [] temp = new float[raster.getWidth() * raster.getHeight() * raster.getNumBands()];
float [] pixels = raster.getPixels(0,0,raster.getWidth(),raster.getHeight(),temp);
for(float pixel: pixels) {
floatList.add(pixel);
}
#记个时
long t = System.currentTimeMillis();
#创建连接,注意usePlaintext设置为true表示用非SSL连接
ManagedChannel channel = ManagedChannelBuilder.forAddress("192.168.2.24", 9000).usePlaintext(true).build();
//这里还是先用block模式
PredictionServiceGrpc.PredictionServiceBlockingStub stub = PredictionServiceGrpc.newBlockingStub(channel);
//创建请求
Predict.PredictRequest.Builder predictRequestBuilder = Predict.PredictRequest.newBuilder();
//模型名称和模型方法名预设
Model.ModelSpec.Builder modelSpecBuilder = Model.ModelSpec.newBuilder();
modelSpecBuilder.setName("classify_data");
modelSpecBuilder.setSignatureName("predict_image");
predictRequestBuilder.setModelSpec(modelSpecBuilder);
//设置入参,访问默认是最新版本,如果需要特定版本可以使用tensorProtoBuilder.setVersionNumber方法
TensorProto.Builder tensorProtoBuilder = TensorProto.newBuilder();
tensorProtoBuilder.setDtype(DataType.DT_FLOAT);
TensorShapeProto.Builder tensorShapeBuilder = TensorShapeProto.newBuilder();
tensorShapeBuilder.addDim(TensorShapeProto.Dim.newBuilder().setSize(1));
#150528 = 224 * 224 * 3
tensorShapeBuilder.addDim(TensorShapeProto.Dim.newBuilder().setSize(150528));
tensorProtoBuilder.setTensorShape(tensorShapeBuilder.build());
tensorProtoBuilder.addAllFloatVal(floatList);
predictRequestBuilder.putInputs("image", tensorProtoBuilder.build());
//访问并获取结果
Predict.PredictResponse predictResponse = stub.predict(predictRequestBuilder.build());
System.out.println("classify is: " + predictResponse.getOutputsOrThrow("classify").getIntVal(0));
System.out.println("prob is: " + predictResponse.getOutputsOrThrow("pro").getFloatVal(0));
System.out.println("cost time: " + (System.currentTimeMillis() - t));
结果打印如下:
classify is: 2
prob is: 1.0
cost time: 6911