我们如何获得 Tensorflow/Tflearn 中隐藏层节点的值？答案

【问题标题】：How can we get the values of hidden layer nodes in Tensorflow/Tflearn?我们如何获得 Tensorflow/Tflearn 中隐藏层节点的值？
【发布时间】：2016-11-12 18:24:18
【问题描述】：

这是 tflearn 中的异或代码。我希望获得倒数第二个隐藏层节点的值（而不是权重）。我怎么能得到那个？更具体地说，我希望为下面给出的四个预测中的每一个获得第 2 层节点的值（在代码中给出）。

import tensorflow as tf
import tflearn

X = [[0., 0.], [0., 1.], [1., 0.], [1., 1.]]  #input
Y_xor = [[0.], [1.], [1.], [0.]]  #input_labels

# Graph definition
with tf.Graph().as_default():
    tnorm = tflearn.initializations.uniform(minval=-1.0, maxval=1.0)
    net = tflearn.input_data(shape=[None, 2], name='inputLayer')
    net = tflearn.fully_connected(net, 2, activation='sigmoid', weights_init=tnorm, name='layer1')
    net = tflearn.fully_connected(net, 1, activation='softmax', weights_init=tnorm, name='layer2')
    regressor = tflearn.regression(net, optimizer='sgd', learning_rate=2., loss='mean_square', name='layer3')

    # Training
    m = tflearn.DNN(regressor)
    m.fit(X, Y_xor, n_epoch=100, snapshot_epoch=False) 

    # Testing
    print("Testing XOR operator")
    print("0 xor 0:", m.predict([[0., 0.]]))
    print("0 xor 1:", m.predict([[0., 1.]]))
    print("1 xor 0:", m.predict([[1., 0.]]))
    print("1 xor 1:", m.predict([[1., 1.]]))

    layer1_var = tflearn.variables.get_layer_variables_by_name('layer1')
    layer2_var = tflearn.variables.get_layer_variables_by_name('layer2')
    inputLayer_var = tflearn.variables.get_layer_variables_by_name('inputLayer')

    #result = tf.matmul(inputLayer_var, layer1_var[0]) + layer1_var[1]

    with m.session.as_default():
        print(tflearn.variables.get_value(layer1_var[0]))   #layer1 weights
        print(tflearn.variables.get_value(layer1_var[1]))   #layer1 bias
        print(tflearn.variables.get_value(layer2_var[0]))   #layer2 weights
        print(tflearn.variables.get_value(layer2_var[1]))   #layer2 bias

【问题讨论】：

标签： tensorflow neural-network tflearn

【解决方案1】：

这可能不能直接回答你的问题，但是如果你使用tflearn，获取每一层的权重就很简单了，

net = tflearn.fully_connected(net, 300)
        self.fc2_w = net.W
        self.fc2_b = net.b

记住一件事，将权重提取代码放在层之后，而不是在批量归一化或单独激活之后

【讨论】：

【解决方案2】：

您可以重复使用共享同一会话的新模型（以使用相同的权重）： .请注意，您也可以保存您的“m”模型并使用“m2”加载它，这会产生类似的结果。

import tensorflow as tf
import tflearn

X = [[0., 0.], [0., 1.], [1., 0.], [1., 1.]]
Y_xor = [[0.], [1.], [1.], [0.]]

# Graph definition
with tf.Graph().as_default():
    tnorm = tflearn.initializations.uniform(minval=-1.0, maxval=1.0)
    net = tflearn.input_data(shape=[None, 2], name='inputLayer')
    layer1 = tflearn.fully_connected(net, 2, activation='sigmoid', weights_init=tnorm, name='layer1')
    layer2 = tflearn.fully_connected(layer1, 1, activation='softmax', weights_init=tnorm, name='layer2')
    regressor = tflearn.regression(layer2, optimizer='sgd', learning_rate=2., loss='mean_square', name='layer3')

    # Training
    m = tflearn.DNN(regressor)
    m.fit(X, Y_xor, n_epoch=100, snapshot_epoch=False) 

    # Testing
    print("Testing XOR operator")
    print("0 xor 0:", m.predict([[0., 0.]]))
    print("0 xor 1:", m.predict([[0., 1.]]))
    print("1 xor 0:", m.predict([[1., 0.]]))
    print("1 xor 1:", m.predict([[1., 1.]]))

    # You can create a new model, that share the same session (to get same weights)
    # Or you can also simply save and load a model
    m2 = tflearn.DNN(layer1, session=m.session)
    print(m2.predict([[0., 0.]]))

【讨论】：