先上代碼

import tensorflow as tf
import numpy as np
from tensorflow.python.platform import gfile
from tensorflow.python.lib.io import file_io

input_tensor_key = 'Placeholder:0'

def loadNpData(filename):
    tensor_key_feed_dict = {}

    #inputs = preprocess_inputs_arg_string(inputs_str)
    data = np.load(file_io.FileIO(filename, mode='r'))

    # When no key is specified for the input file.
    # Check if npz file only contains a single numpy ndarray.
    if isinstance(data, np.lib.npyio.NpzFile):
        variable_name_list = data.files
        if len(variable_name_list) != 1:
            raise RuntimeError(
                'Input file %s contains more than one ndarrays. Please specify '
                'the name of ndarray to use.' % filename)
        tensor_key_feed_dict[input_tensor_key] = data[variable_name_list[0]]
    else:
        tensor_key_feed_dict[input_tensor_key] = data
    return tensor_key_feed_dict

with tf.Session() as sess:
    # 定義模型文件及樣本測(cè)試文件
    model_filename = 'merge1_graph.pb'
    example_png = 'examples.npy'
    # 加載npy格式的圖片測(cè)試樣本數(shù)據(jù)
    image_data = loadNpData(example_png)
    #加載模型文件
    with gfile.FastGFile(model_filename, 'rb') as f:
        graph_def = tf.GraphDef();
        graph_def.ParseFromString(f.read())

    # 獲取輸入節(jié)點(diǎn)的tensor
    inputs = sess.graph.get_tensor_by_name("Placeholder:0");
    #打印輸入節(jié)點(diǎn)的信息
    #print inputs
    # 導(dǎo)入計(jì)算圖证杭，定義輸入節(jié)點(diǎn)及輸出節(jié)點(diǎn)
    output = tf.import_graph_def(graph_def, input_map={'Placeholder:0':inputs}, return_elements=[ 'ArgMax:0','Softmax:0']) 
    # 打印輸出節(jié)點(diǎn)的信息
    #print output
    results = sess.run(output, feed_dict={inputs:image_data[input_tensor_key]})
    print 'ArgMax result(預(yù)測(cè)結(jié)果對(duì)應(yīng)的標(biāo)簽值):'  
    print results[0]
    print 'Softmax result(最后一層的輸出):'
    print results[1]
    # 輸出node詳細(xì)信息，此處默認(rèn)只打印第一個(gè)節(jié)點(diǎn)
    for node in graph_def.node:
        print node
        break

運(yùn)行輸出

ArgMax result(預(yù)測(cè)結(jié)果對(duì)應(yīng)的標(biāo)簽值):
[3 3]
Softmax result(最后一層的輸出):
[[4.1668140e-12 9.0696268e-18 6.4261091e-13 9.9999940e-01 1.7161388e-30
  5.4321697e-07 7.6357951e-09 6.3293229e-19 1.3812791e-13 1.5360580e-12]
 [1.1472046e-05 3.3404951e-10 6.0365837e-09 9.9997592e-01 9.8635665e-15
  5.7557719e-07 1.1977763e-05 1.6275100e-16 7.2288098e-10 5.0601763e-08]]

此處加載的關(guān)鍵在于tf.import_graph_def函數(shù)的參數(shù)配置俭令，三個(gè)參數(shù)graph_def input_map return_elements

第一個(gè)參數(shù)是導(dǎo)入的圖
input_map是指定輸入節(jié)點(diǎn)拂封，如果不指定换吧，后面run的時(shí)候會(huì)報(bào)錯(cuò) ==You must feed a value for placeholder tensor 'Placeholder'==

return_elements 是指定運(yùn)算后的輸出節(jié)點(diǎn)泌霍，此處就是我們想要得到的標(biāo)簽估計(jì)值 ArgMax 以及 最后一層節(jié)點(diǎn)輸出 Softmax

模型的測(cè)試參考 將Tensorflow模型導(dǎo)出為單個(gè)文件