Tensorflow Serving实战 保存自己训练的模型到SavedModel并serving（二）

开发环境

• Ubuntu 18.x
• Docker 18.x
• Tensorflow 1.x

构建模型signature

"""
tf.saver.restore(...)
"""
# 上面省略了模型的加载的代码

#设置导出的路径
export_path_base = "serving"
export_version = '1';
export_path = os.path.join(export_path_base,export_version)
print('Exporting trained model to', export_path)
builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(export_path)
print('step1 => builder  created successfully')
# 构建signature_def_map，参数为模型的输入与输出
tensor_info_input = tf.compat.v1.saved_model.utils.build_tensor_info(self.model.input)
tensor_info_output = tf.compat.v1.saved_model.utils.build_tensor_info(self.model.out)
# 定义模型的类型
prediction_signature = (
            tf.compat.v1.saved_model.signature_def_utils.build_signature_def(
                inputs={'images': tensor_info_input},
                outputs={'result': tensor_info_output},
                method_name=tf.saved_model.PREDICT_METHOD_NAME))
print('step2 => prediction_signature created successfully')
builder.add_meta_graph_and_variables(
            # tags:SERVING,TRAINING,EVAL,GPU,TPU
            self.sess, [tf.saved_model.SERVING],
            signature_def_map={'predict_images': prediction_signature,
            })
print('step3 => builder successfully add meta graph and variables\nNext is to export model...')
#builder.save(as_text=True)
builder.save()
print('Done exporting!')

导出的文件结构

├── serving                         //模型文件 
│   ├──1                            //版本号
│   │   ├── saved_model.pb                
│   │   ├── variables
│   │   │   ├── variables
│   │   │   ├── variables.data-00000-of-00001          
│   │   │   ├── variables.index        

用docker Serving自己的模型

略

使用pyhon的requests调用模型

import requests
import skimage.io as skio
import numpy as np
import matplotlib.pyplot as plt

SERVER_URL = 'http://*.*.*.*:8501/v1/models/super-resolution:predict'
class Serving(object):
    def __init__(self):
        self.img_size = 48
        self.scale = 3
        
    def main(self):
        # 图片路径，只能是灰度图像
        img = skio.imread('./testImage/Set5/butterfly_GT.bmp', as_gray=True);
        img_narray = np.asarray(img)
        img_narray = img_narray[0:48,0:48]

        # 模型能接受的图片大小只能是[48,48],且格式为[1,48,48,1]
        img_narray = img_narray.reshape([1,img_narray.shape[0],img_narray.shape[1], 1])
        
        #构建请求数据，这里的格式很容易错，如果错了,会说：400 Client Error: Bad Request for url:****
        predict_request = '{"signature_name": "predict_images", "instances":%s }' % img_narray.tolist()
        response = requests.post(SERVER_URL, data=predict_request)
        response.raise_for_status()

        #获取预测结果
        prediction = response.json()['predictions']
        prediction_img = np.array(prediction)
        plt.imshow(prediction_img.reshape([144, 144]), 'gray')
        plt.show()

if __name__ == '__main__':
    P_get = requests.get("http://101.37.14.171:8501/v1/models/super-resolution")
    print(P_get.content)
    s = Serving()
    s.main()

补充：使用keras导出.pb格式的模型

keras原本的模型保存格式是h5格式的，

# 导入加载模型的方法load_model和模型编译使用的优化器类Adam
from keras.models import load_model
from keras.optimizers import Adam
import tensorflow as tf
from keras import backend as K
import os
# 加载已经训练好的ResNet50图片分类模型的模型文件
model_filePath = './my_model_resnet.h5'
model = load_model(model_filePath)
model.compile(loss='categorical_crossentropy',
              optimizer=Adam(lr=0.001),
              metrics=['accuracy'])
#模型导出到Tensorflow Serving的代码
export_path_base = "serving"
export_version = '1';
export_path = os.path.join(export_path_base,
                           export_version)
print('Exporting trained model to', export_path)
builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(export_path)
# Build the signature_def_map.
tensor_info_input = tf.compat.v1.saved_model.utils.build_tensor_info(model.input)
tensor_info_output = tf.compat.v1.saved_model.utils.build_tensor_info(model.output)

# 定义模型的类型
prediction_signature = (
    tf.compat.v1.saved_model.signature_def_utils.build_signature_def(
        inputs={'images': tensor_info_input},
        outputs={'result': tensor_info_output},
        method_name=tf.saved_model.PREDICT_METHOD_NAME))
print('step3 => prediction_signature created successfully')
builder.add_meta_graph_and_variables(
    # tags:SERVING,TRAINING,EVAL,GPU,TPU
    K.get_session(), [tf.saved_model.SERVING],
    signature_def_map={
        'predict_images': prediction_signature,
    })
print('step4 => builder successfully add meta graph and variables\nNext is to export model...')
#builder.save(as_text=True)
builder.save()
print('Done exporting!')