在 tensorflow 中使用两种不同的模型

Question

我正在尝试使用两种不同的 mobilenet 模型。以下是我如何初始化模型的代码。

def initialSetup():
    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
    start_time = timeit.default_timer()

    # This takes 2-5 seconds to run
    # Unpersists graph from file
    with tf.gfile.FastGFile('age/output_graph.pb', 'rb') as f:
        age_graph_def = tf.GraphDef()
        age_graph_def.ParseFromString(f.read())
        tf.import_graph_def(age_graph_def, name='')

    with tf.gfile.FastGFile('output_graph.pb', 'rb') as f:
        gender_graph_def = tf.GraphDef()
        gender_graph_def.ParseFromString(f.read())
        tf.import_graph_def(gender_graph_def, name='')

    print ('Took {} seconds to unpersist the graph'.format(timeit.default_timer() - start_time))

由于两者都是两个不同的模型，我该如何使用它进行预测？

更新

initialSetup()

age_session = tf.Session(graph=age_graph_def)
gender_session = tf.Session(graph=gender_graph_def)

with tf.Session() as sess:
    start_time = timeit.default_timer()

    # Feed the image_data as input to the graph and get first prediction
    softmax_tensor = age_session.graph.get_tensor_by_name('final_result:0')

    print ('Took {} seconds to feed data to graph'.format(timeit.default_timer() - start_time))

    while True:
        # Capture frame-by-frame
        ret, frame = video_capture.read()

错误

Traceback（最近一次调用最后一次）：文件 “C:/Users/Desktop/untitled/testimg/testimg/combo.py”，第 48 行，在 age_session = tf.Session(graph=age_graph_def) 文件 "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\client\session.py", 第 1292 行，在 init 中 super(Session, self).init(target, graph, config=config) File "C:\Program 文件\Anaconda3\lib\site-packages\tensorflow\python\client\session.py", 第 529 行，在 init 中 raise TypeError('graph must be a tf.Graph, but got %s' % type(graph)) TypeError: graph must be a tf.Graph, but got Exception ignored in: > Traceback（最近一次调用最后一次）：文件 "C:\程序 文件\Anaconda3\lib\site-packages\tensorflow\python\client\session.py", 第 587 行，在 del 如果 self._session 不是 None: AttributeError: 'Session' object has no attribute '_session'

Answer 1

当您在同一个图中使用多个模型时，使用名称范围为各个张量提供可预测的名称。例如，您可以将initial_setup() 改写如下：

def initialSetup():
    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
    start_time = timeit.default_timer()

    # This takes 2-5 seconds to run
    # Unpersists graph from file
    with tf.gfile.FastGFile('age/output_graph.pb', 'rb') as f:
        age_graph_def = tf.GraphDef()
        age_graph_def.ParseFromString(f.read())
        tf.import_graph_def(age_graph_def, name='age_model')

    with tf.gfile.FastGFile('output_graph.pb', 'rb') as f:
        gender_graph_def = tf.GraphDef()
        gender_graph_def.ParseFromString(f.read())
        tf.import_graph_def(gender_graph_def, name='gender_model')

    print ('Took {} seconds to unpersist the graph'.format(timeit.default_timer() - start_time))

现在age_graph_def 中的所有节点的名称都将以"age_model/" 为前缀，gender_graph_def 中的所有节点的名称都将以"gender_model/" 为前缀。它们都是同一个默认图的一部分，因此您可以使用单个 tf.Session 而不使用 graph 参数来访问任一模型。

initialSetup()

with tf.Session() as sess:
    start_time = timeit.default_timer()

    # Feed the image_data as input to the graph and get first prediction
    softmax_tensor = sess.graph.get_tensor_by_name('age_model/final_result:0')

    # Alternatively, to get a tensor from the gender model:
    # tensor = sess.graph.get_tensor_by_name('gender_model/...')

    print ('Took {} seconds to feed data to graph'.format(timeit.default_timer() - start_time))

    while True:
        # Capture frame-by-frame
        ret, frame = video_capture.read()

Answer 2

tf.Session 需要 tf.Graph 实例而不是 tf.GraphDef，请按照以下步骤解决问题。

def initialSetup():
    with tf.gfile.FastGFile('age/output_graph.pb', 'rb') as f:
        age_graph_def = tf.GraphDef()
        age_graph_def.ParseFromString(f.read())
        with tf.Graph().as_default() as graph:
            tf.import_graph_def(age_graph_def, name='')
            age_graph = graph

   ...
   return age_graph, gender_graph

age_graph, gender_graph = initial_setup() 
age_session = tf.Session(graph=age_graph)
...
# also delete the following line, as it creates another new context 
with tf.Session() as sess: