AI實戰：使用pb模型做推理

模型應用

模型部署應用首選Tensorflow，Tensorflow模型部署使用pb格式最爲簡單。

本文以圖像分類模型爲例，介紹pb模型使用方法：

• 代碼

import cv2
import tensorflow as tf
import numpy as np
import sys, os


class Recognizer():
    def __init__(self, pb_path):
        
        self.pb_path = pb_path
        self.config = tf.ConfigProto()
        self.config.gpu_options.allow_growth = True
        self.init_model()
        
        
    def init_model(self):
    
        tf.Graph().as_default()
        self.output_graph_def = tf.GraphDef()
        with open(self.pb_path, 'rb') as f:
            self.output_graph_def.ParseFromString(f.read())
            tf.import_graph_def(
                                self.output_graph_def,
                                input_map = None,
                                return_elements = None,
                                name = None,
                                op_dict = None,
                                producer_op_list = None
                            )
                            
        self.sess = tf.Session(config = self.config)
        self.input = self.sess.graph.get_tensor_by_name("input_1:0")#自己定義的輸入tensor名稱
        self.output = self.sess.graph.get_tensor_by_name("output_1:0")#自己定義的輸出tensor名稱
        
        
    def predict(self, img):
        
        img = (img - 255/2.0) / 255
        img = img[np.newaxis, :, :, :]
        res = self.sess.run(self.output, feed_dict={self.input: img})
        class_id = np.argmax(res)
        
        return str(class_id)
        
    def batch_predict(self, img_list):
        
        class_ids = []
        for img in img_list:
            class_id = self.predict(img)
            class_ids.append(class_id)
            
        return class_ids
        
        
if __name__ == '__main__':

    if len(sys.argv) == 3:
        recognizer = Recognizer(pb_path=sys.argv[1])
        img = cv2.imread(sys.argv[2])
        res = recognizer.predict(img)
        print('result:', res)

注意：
1、input、output的tensor名稱是網絡中自己定義的名稱，未定義則默認爲inut_1:0、output_1:0

推薦：

AI實戰：查看pb模型的graph的所有層的名稱