1

所以我创建了自己的神经网络，我想对输入变量进行自动微分。我的神经网络代码是这样的

n_input = 1     
n_hidden_1 = 50 
n_hidden_2 = 50 
n_output = 1 

weights = {
'h1': tf.Variable(tf.random.normal([n_input, n_hidden_1],0,0.5)),
'h2': tf.Variable(tf.random.normal([n_hidden_1, n_hidden_2],0,0.5)),
'out': tf.Variable(tf.random.normal([n_hidden_2, n_output],0,0.5))
}

biases = {
'b1': tf.Variable(tf.random.normal([n_hidden_1],0,0.5)),
'b2': tf.Variable(tf.random.normal([n_hidden_2],0,0.5)),
'out': tf.Variable(tf.random.normal([n_output],0,0.5))
}

def multilayer_perceptron(x):
    x = np.array([[[x]]],  dtype='float32')
    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])
    layer_1 = tf.nn.tanh(layer_1)
    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])
    layer_2 = tf.nn.tanh(layer_2)
    output = tf.matmul(layer_2, weights['out']) + biases['out']
    return output

并且tf.GradientTape()，我试图用这个来区分神经网络

x = tf.Variable(1.0)
with tf.GradientTape() as tape:
    y = multilayer_perceptron(x)
dNN1 = tape.gradient(y,x)
print(dNN1)

哪个结果None。我在这里做错了什么？

4

2 回答 2

1

因为您正在转换为不可区分x的 numpy 数组。np.array

像这样修改你的代码：

def multilayer_perceptron(x):
    #x = np.array([[[x]]],  dtype='float32') #comment
    layer_1 = tf.add(tf.matmul([[[x]]], weights['h1']), biases['b1']) #change x shape by adding []
    layer_1 = tf.nn.tanh(layer_1)
    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])
    layer_2 = tf.nn.tanh(layer_2)
    output = tf.matmul(layer_2, weights['out']) + biases['out']
    return output
于 2021-07-24T16:44:20.123 回答
0

为了良好运行一些 tensorflow 操作，最好所有操作元素都是 tf.tensor 类型，您必须使用重塑

def multilayer_perceptron(x):
 x =  tf.reshape(x , (1,1,1))
于 2021-07-24T17:12:00.130 回答