我的tensorflow模型在一个类中定义如下：
class EENetwork:
def __init__(self, state_size, learning_rate, output_units, name='EENetwork'):
self.learning_rate = learning_rate
self.state_size = state_size # State size = height * width * cameras
self.output_units = output_units
with tf.variable_scope(name):
self.inputs_ = tf.placeholder(tf.float32, [None, *state_size], name='inputs')
self.outputs_ = tf.placeholder(tf.float32, [None, self.output_units], name='outputs_') # To hold the training Y
# Placeholders aren't added to the saved graph by default
tf.add_to_collection('inputs', self.inputs_)
tf.add_to_collection('outputs_', self.outputs_)
# Conv1 = samples * h * w * 24
# Assume h, w = 64, so samples * 64 * 64 * 24
self.conv1 = tf.layers.conv2d(inputs=self.inputs_, filters=24, kernel_size=[5, 5],
padding='same', activation=tf.nn.relu, name='conv1')
# Maxpool1 = samples * h/2 * w/2 * 12 = samples * 32 * 32 * 24
self.maxpool1 = tf.layers.max_pooling2d(inputs=self.conv1, pool_size=[2, 2], strides=2, name='mp1')
# Conv2 = samples * 32 * 32* 36
self.conv2 = tf.layers.conv2d(inputs=self.maxpool1, filters=36, kernel_size=[5, 5],
padding='same', activation=tf.nn.relu, name='conv2')
# Maxpool2 = samples * 16 * 16 * 36
self.maxpool2 = tf.layers.max_pooling2d(inputs=self.conv2, pool_size=[2, 2], strides=2, name='mp2')
# Conv3 = samples * 16 * 16 * 48
self.conv3 = tf.layers.conv2d(inputs=self.maxpool2, filters=48, kernel_size=[3, 3],
padding='same', activation=tf.nn.relu, name='conv3')
# Maxpool3 = samples * 8 * 8 * 48
self.maxpool3 = tf.layers.max_pooling2d(inputs=self.conv3, pool_size=[2, 2], strides=2, name='mp3')
# Conv4 = samples * 8 * 8 * 64
self.conv4 = tf.layers.conv2d(inputs=self.maxpool3, filters=64, kernel_size=[3, 3],
padding='same', activation=tf.nn.relu, name='conv4')
# Maxpool4 = samples * 4 * 4 * 64
self.maxpool4 = tf.layers.max_pooling2d(inputs=self.conv4, pool_size=[2, 2], strides=2, name='mp4')
# Flatten = samples * 1024
self.flatten = tf.contrib.layers.flatten(self.maxpool4)
# FC1, Output units = 1164
self.fc1 = tf.layers.dense(inputs=self.flatten, units=1164, activation=tf.nn.elu,
kernel_initializer=tf.contrib.layers.xavier_initializer(), name='fc1')
# FC2, Output units = 100
self.fc2 = tf.layers.dense(inputs=self.fc1, units=100, activation=tf.nn.elu,
kernel_initializer=tf.contrib.layers.xavier_initializer(), name='fc2')
# FC3, Output units = 50
self.fc3 = tf.layers.dense(inputs=self.fc2, units=50, activation=tf.nn.elu,
kernel_initializer=tf.contrib.layers.xavier_initializer(), name='fc3')
# FC4, Output units = 10
self.fc4 = tf.layers.dense(inputs=self.fc3, units=10, activation=tf.nn.elu,
kernel_initializer=tf.contrib.layers.xavier_initializer(), name='fc4')
# Output
self.output = tf.layers.dense(inputs=self.flatten, units=self.output_units, name='DL')
# Cost
#self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=self.output, labels=self.outputs_))
self.loss = tf.losses.mean_squared_error(labels=self.outputs_, predictions=self.output)
# Optimizer
self.optimizer = tf.train.AdamOptimizer(self.learning_rate).minimize(self.loss)
我正在创建此类的一个实例，然后对其进行培训和保存。现在，将保存的模型还原到另一个文件中之后，我想获得self.output产生的输出。
这是我为第二个文件编写的代码：
with tf.Session() as sess:
saver = tf.train.import_meta_graph("./models/drivingModel_{LR}_{EP}_{MN}.ckpt.meta".format(LR=learning_rate, EP=epochs, MN=mod_number))
saver.restore(sess, "./models/drivingModel_{LR}_{EP}_{MN}.ckpt".format(LR=learning_rate, EP=epochs, MN=mod_number))
frontImage = cv2.imread('Front_5.png')
leftImage = cv2.imread('Left_5.png')
rightImage = cv2.imread('Right_5.png')
front = cv2.resize(cv2.cvtColor(frontImage, cv2.COLOR_RGB2GRAY), (h, w))
left = cv2.resize(cv2.cvtColor(leftImage, cv2.COLOR_RGB2GRAY), (h, w))
right = cv2.resize(cv2.cvtColor(rightImage, cv2.COLOR_RGB2GRAY), (h, w))
images = np.stack((front, left, right)) # num_cameras * h * w
images = np.moveaxis(images, [0, 1, 2], [2, 0, 1])
X = []
X.append(images)
X = np.array(X)
print(tf.get_default_graph().get_all_collection_keys())
all_variables = tf.get_collection('variables')
print("Variables of the tensor:")
for var in all_variables:
print(var)
graph = sess.graph
print("Operations of the tensor: ")
operations = tf.get_collection('train_op')
for op in graph.get_operations():
print(op)
我想将X作为feeddict参数传递给self.inputs_并获取self.output产生的输出（命名为“ DL”）。
我不知道我可以使用哪些变量和操作来执行上述任务。
我可以使用以下方法获取输入占位符：
inputs_ = tf.get_collection('inputs')[0]
但是如何将X传递给它并使用它来获取self.output层产生的值？
编辑： 我可以通过创建该类的新实例然后运行以下代码来解决该问题：
tf.reset_default_graph()
drivingNet = cnn_model.EENetwork(state_size=state_size,
learning_rate=learning_rate, output_units=classes)
with tf.Session() as sess:
saver = tf.train.Saver(tf.global_variables())
saver = tf.train.import_meta_graph("./models/drivingModel_{LR}_{EP}_{MN}.ckpt.meta".format(LR=learning_rate, EP=epochs, MN=mod_number))
saver.restore(sess, "./models/drivingModel_{LR}_{EP}_{MN}.ckpt".format(LR=learning_rate, EP=epochs, MN=mod_number))
frontImage = cv2.imread('Front_5.png')
leftImage = cv2.imread('Left_5.png')
rightImage = cv2.imread('Right_5.png')
front = cv2.resize(cv2.cvtColor(frontImage, cv2.COLOR_RGB2GRAY), (h, w))
left = cv2.resize(cv2.cvtColor(leftImage, cv2.COLOR_RGB2GRAY), (h, w))
right = cv2.resize(cv2.cvtColor(rightImage, cv2.COLOR_RGB2GRAY), (h, w))
images = np.stack((front, left, right)) # num_cameras * h * w
# Move the axis to make it h * w * num_cameras
images = np.moveaxis(images, [0, 1, 2], [2, 0, 1])
# images = np.concatenate((leftImage, frontImage), axis=1)
# images = np.concatenate((images, rightImage), axis=1)
# Input needs to be passed as samples * h * w * num_cameras
X = []
X.append(images)
X = np.array(X) # This makes it 1 * h * w * cameras
decision = sess.run(drivingNet.output, feed_dict={drivingNet.inputs_: X})
temp = decision.tolist()
flat_list = [item for sublist in temp for item in sublist]
print(flat_list)
但是，在这种情况下，我仍然不知道saver = tf.train.Saver（tf.global_variables（））在做什么？为什么需要此行代码？如果我不包括这一行，它将引发“试图使用未初始化的值”错误。元图和saver.restore不应该将所有值初始化为保存的值吗？