Question

在下面的代码中，test是一个3维的numpy数组，阵列中的矩阵具有不同的行大小。因为这个原因，我无法在feed dict中提供这个3-d数组。有什么方法可以解决这个问题吗？

这是错误：

Traceback (most recent call last):
  File "test.py", line 
83, in <module>
    sess.run(iter.initializer, feed_dict={M: test, v_a: va, y: label})
  File "C:\WinPython-64bit-3.6.3.0Qt5\python-3.6.3.amd64\lib\site- 
packages\tensorflow\python\client\session.py", line 905, in run
    run_metadata_ptr)
  File "C:\WinPython-64bit-3.6.3.0Qt5\python-3.6.3.amd64\lib\site- 
packages\tensorflow\python\client\session.py", line 1106, in _run
    np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)
  File "C:\WinPython-64bit-3.6.3.0Qt5\python-3.6.3.amd64\lib\site- 
packages\numpy\core\numeric.py", line 531, in asarray
    return array(a, dtype, copy=False, order=order)
ValueError: setting an array element with a sequence.

这是我的代码：

def multilayer_perceptron(m_i, v_a, weights, biases):

    #Hidden layers

    layer_1 = tf.add(tf.add(tf.matmul(weights['W1'], m_i), 
    tf.matmul(weights['W2'], v_a)), biases['b1'])
    layer_1 = tf.nn.tanh(layer_1)

    layer_2 = tf.matmul(weights['W3'], layer_1)

    return layer_2

def ModelA(M, v_a, weights, biases, d):

    N = sess.run(tf.size(M))/d

    c = multilayer_perceptron((tf.slice(M, [0, 0], ([d, 1]))), v_a, weights, 
     biases)

    for i in range(1, N):

        c = tf.concat([c, multilayer_perceptron((tf.slice(M, [0, i], [d, 
            1])), v_a, weights, biases)], axis=0)

    alpha = tf.nn.softmax(tf.reshape(c, [-1]))

    v_ns = tf.matmul(M, tf.reshape(alpha, [N, 1]))

    layer_3 = tf.add(tf.matmul(weights['W4'], v_ns), biases['b2'])
    v_ms = tf.nn.tanh(layer_3)

    layer_4 = tf.add(tf.matmul(weights['W5'], v_ms), biases['b3'])
    pred = tf.nn.softmax(tf.reshape(layer_4, [-1]))

    return pred

def generator():
    for el in test:
        yield el

# Placeholders and Constants
number_of_classes = 3
M = tf.placeholder(tf.float32, name='M')
d = 50     #rows M matrix = d
y = tf.placeholder(tf.float32, shape=[1, number_of_classes], name='y')
v_a = tf.placeholder(tf.float32, shape=[d, 1], name='v_a')

# Hyperparameters
learning_rate = 0.01
training_epochs = 10

# Variables
weights = {
    'W1': tf.Variable(tf.truncated_normal(shape=[d, d])),
    'W2': tf.Variable(tf.truncated_normal(shape=[d, d])),
    'W3': tf.Variable(tf.truncated_normal(shape=[1, d])),
    'W4': tf.Variable(tf.truncated_normal(shape=[d, d])),
    'W5': tf.Variable(tf.truncated_normal(shape=[number_of_classes, d]))
}

biases = {
    'b1': tf.Variable(tf.truncated_normal([d, 1])),
    'b2': tf.Variable(tf.truncated_normal([d, 1])),
    'b3': tf.Variable(tf.truncated_normal([number_of_classes, 1]))
}

init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)

test = np.array([[[[1.], [2.], [3.]],[[4.], [5.], [6.]], [[7.], [8.], 
        [9.]]], [[[6.], [2.], [4.]], [[2.], [1.], [8.]]], [[[7.], [6.], 
        [2.]], [[2.], [4.], [2.]], [[3.], [8.], [9.]], [[1.], [2.], [1.]]]])
label = np.array([[0, 1, 0], [0, 0, 1], [1, 0, 0]])
va = np.array([[[1.], [2.], [3.]], [[2.], [1.], [8.]], [[1.], [2.], [1.]]])


dataset = tf.data.Dataset().from_generator(generator, 
output_types=tf.float32)
iter = dataset.make_initializable_iterator()

sess.run(iter.initializer, feed_dict={M: test, v_a: va, y: label})
cost_function = 
    tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=ModelA(M, 
        v_a, weights, biases, d), labels=label))
optimizer = tf.train.GradientDescentOptimizer(0.01).minimize(cost_function)

print('Training...')
for i in range(training_epochs):
    tot_loss = 0
    trainings, loss_value = sess.run([optimizer, cost_function])
    tot_loss += loss_value
    print("Iter: {}, Loss: {:.4f}".format(i, tot_loss))

Answer 1

当你还在使用tf.operation时，不要使用sess.run（...），这会使你的代码慢得多。可以通过用零填充矩阵并使其成为3D来解决变化的行大小。

通过python中的feed-dict提供数据时出错

1 个答案: