tensorflow serving (三):request/g
2019-05-09 本文已影响0人
夕一啊
保存pb模型的时候就定义好了输入输出节点的名字,这里是input和output
request
使用docker起服务,默认的8501端口是request端口
传入数据是 图片,json不接受np array, 所以需要转成list,但是这样很慢
import os
import requests
from time import time
import cv2
import json
import numpy as np
url = 'http://localhost:7475/v1/models/pse:predict'
def predict_one(img_path):
start_time = time()
print(img_path)
img = cv2.imread(img_path)
time1 = time()
print("read pic", time1 - start_time, " s")
# predict_request_data = '{"instances" : [{"input": %s}]}' % img.tolist() # json格式对应json, data格式对应字符串
predict_request_json = {"instances" : [{"input": img.tolist() }]}
time2 = time()
print("to list", time2 - time1, " s")
r = requests.post(url, json=predict_request_json)
#r = requests.post(url, data=predict_request_data) # json格式对应json, data格式对应字符串
kernel = np.array(json.loads(r.content.decode("utf-8"))["predictions"][0])[:,:,0]
print(kernel.shape)
end_time = time()
print(end_time - start_time, " s")
gRPC
gRPC的端口映射到宿主机的8500端口,不需要转换图片为list速度会快很多
pip install tensorflow_serving_api
import tensorflow as tf
from tensorflow_serving.apis import predict_pb2
from tensorflow_serving.apis import prediction_service_pb2_grpc
import grpc
import cv2
import numpy as np
from time import time
import os
channel = grpc.insecure_channel("localhost:7475")
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
request = predict_pb2.PredictRequest()
# 指定启动tensorflow serving时配置的model_name和是保存模型时的方法名
request.model_spec.name = "pse"
request.model_spec.signature_name = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
def predct_one(pic_path):
print(pic_path)
start_time = time()
img = cv2.imread(pic_path)
print("orig size:", img.shape)
# img = cv2.resize(img,(1280,960))
img = np.expand_dims(img, 0)
# 输入tensor后序列化
request.inputs["input"].ParseFromString(tf.contrib.util.make_tensor_proto(img, dtype=tf.float32).SerializeToString())
time1 = time()
print("read pic", time1 - start_time, " s")
response = stub.Predict(request)
results = {}
for key in response.outputs:
tensor_proto = response.outputs[key]
results[key] = tf.contrib.util.make_ndarray(tensor_proto)
# 从results中取所需要的结果
kernel = results["output"]
print(kernel.shape)
print(time() - start_time, " s")