Question

假设我使用Keras定义了一个网络，如下所示：

model = Sequential()
model.add(Dense(10, input_shape=(10,), activation='relu'))
model.add(Dense(10, activation='relu'))
model.add(Dense(9, activation='sigmoid'))

它有一些重量：

[array([[ 0.33494413, -0.34308964,  0.12796348,  0.17187083, -0.40254939,
    -0.06909397, -0.30374748,  0.14217842,  0.41163749, -0.15252003],
   [-0.07053435,  0.53712451, -0.43015254, -0.28653857,  0.53299475, ...

当我给它一些输入时：

[[ 0.   0.5  0.   0.5  1.   1.   0.   0.5  0.5  0.5]]

它会产生一些输出：

[0.5476531982421875, 0.5172237753868103, 0.5247090458869934, 0.49434927105903625, 0.4599153697490692, 0.44612908363342285, 0.4727349579334259, 0.5116984844207764, 0.49565717577934265]

而期望的输出是：

[0.6776225034927386, 0.0, 0.5247090458869934, 0.0, 0.0, 0.0, 0.4727349579334259, 0.5116984844207764, 0.49565717577934265]

制作错误值：

[0.12996930525055106, -0.5172237753868103, 0.0, -0.49434927105903625, -0.4599153697490692, -0.44612908363342285, 0.0, 0.0, 0.0]

然后我可以按如下方式计算评估的梯度：

outputTensor = model.output
listOfVariableTensors = model.trainable_weights
gradients = k.gradients(outputTensor, listOfVariableTensors)
trainingInputs = inputs
sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())
evaluated_gradients = sess.run(gradients, feed_dict={model.input: trainingInputs})

评估渐变的结果：

[array([[ 0.        ,  0.        ,  0.        ,  0.        ,  0.        ,
     0.        ,  0.        ,  0.        ,  0.        ,  0.        ],
   [ 0.        ,  0.        ,  0.        ,  0.        ,  0.        ,
     0.        ,  0.        ,  0.        ,  0.        ,  0.        ],
   [ 0.01015381,  0.        ,  0.        ,  0.03375177, -0.05576257,
     0.03318337, -0.02608909, -0.06644543, -0.03461133,  0.        ],
   [ 0.02030762,  0.        ,  0.        ,  0.06750354, -0.11152515,
     0.06636675, -0.05217818, -0.13289087, -0.06922265,  0.        ],...

我想用这些渐变来调整我的模型，但我正在失去对数学和数学的追踪。反向传播理论。我是在正确的轨道上吗？

Keras反向传播

0 个答案: