tensorflow报错:Shape must be rank 2 but is rank 3 for 'MatMul' (op: 'MatMul')

tensorflow矩阵相乘，秩不同报错

在tensorflow中写了这样一句：

[python] view plain copy print?y_out = tf.matmul(outputs, W)

y_out = tf.matmul(outputs, W)

其中，outputs的shape为[16,336,400]，W的shape为[400,
4000
1]

出现以下报错：

Shape must be rank 2 but is rank 3 for ‘MatMul’ (op: ‘MatMul’) with input shapes: [16,336,400], [400,1].

Numpy下同样的写法没有问题

[python] view plain copy print?import numpy as np

A = np.array([[[1, 2, 3, 4],
[5, 6, 7, 8],
[9, 0, 1, 2]],
[[4, 3, 2, 1],
[8, 7, 6, 5],
[2, 1, 0, 9]]])
print(A)
print(A.shape)
print(‘—————————’)

B = np.array([[1], [2], [3], [4]])
print(B)
print(B.shape)
print(‘—————————’)

C = np.matmul(A, B)
print(C)
print(C.shape)

import numpy as np

A = np.array([[[1, 2, 3, 4],
[5, 6, 7, 8],
[9, 0, 1, 2]],
[[4, 3, 2, 1],
[8, 7, 6, 5],
[2, 1, 0, 9]]])
print(A)
print(A.shape)
print('---------------------------')

B = np.array([[1], [2], [3], [4]])
print(B)
print(B.shape)
print('---------------------------')

C = np.matmul(A, B)
print(C)
print(C.shape)

输出结果：

[python] view plain copy print?[[[1 2 3 4]
[5 6 7 8]
[9 0 1 2]]

[[4 3 2 1]
[8 7 6 5]
[2 1 0 9]]]
(2, 3, 4)
—————————
[[1]
[2]
[3]
[4]]
(4, 1)
—————————
[[[30]
[70]
[20]]

[[20]
[60]
[40]]]
(2, 3, 1)

[[[1 2 3 4]
[5 6 7 8]
[9 0 1 2]]

[[4 3 2 1]
[8 7 6 5]
[2 1 0 9]]]

(2, 3, 4)
---------------------------
[[1]
[2]
[3]
[4]]

(4, 1)
---------------------------
[[[30]
[70]
[20]]

[[20]
[60]
[40]]]
(2, 3, 1)

解决办法

[python] view plain copy print?import numpy as np import tensorflow as tf sess = tf.Session() A = np.array([[[1, 2, 3, 4], [5, 6, 7, 8], [9, 0, 1, 2]], [[4, 3, 2, 1], [8, 7, 6, 5], [2, 1, 0, 9]]]) B = np.array([[1], [2], [3], [4]]) A = tf.cast(tf.convert_to_tensor(A), tf.int32) # shape=[2, 3, 4] B = tf.cast(tf.convert_to_tensor(B), tf.int32) # shape=[4, 1]#—————————————–修改部分（开始）—————————————–
#要想让A和B进行tf.matmul操作，第一个维数必须一致。因此要把B先tile后转成[2, 4, 1]维
B_ = tf.tile(B, [2, 1])# B的第一维复制2倍，第二维复制1倍
B = tf.reshape(B_, [2, 4, 1])
# 或 更通用的改法：
#B_ = tf.tile(B, [tf.shape(A)[0], 1])
#B = tf.reshape(B_, [tf.shape(A)[0], tf.shape(B)[0], tf.shape(B)[1]])
#—————————————–修改部分（结束）—————————————–

#此时就可以matmul了
C = tf.matmul(A, B)
print(‘C:’,C.get_shape().as_list())
sess.run(C)

import numpy as np
import tensorflow as tf
sess = tf.Session()

A = np.array([[[1, 2, 3, 4],
[5, 6, 7, 8],
[9, 0, 1, 2]],
[[4, 3, 2, 1],
[8, 7, 6, 5],
[2, 1, 0, 9]]])
B = np.array([[1], [2], [3], [4]])

A = tf.cast(tf.convert_to_tensor(A), tf.int32) # shape=[2, 3, 4]
B = tf.cast(tf.convert_to_tensor(B), tf.int32) # shape=[4, 1]

#-----------------------------------------修改部分（开始）-----------------------------------------

#要想让A和B进行tf.matmul操作，第一个维数必须一致。因此要把B先tile后转成[2, 4, 1]维
B_ = tf.tile(B, [2, 1])# B的第一维复制2倍，第二维复制1倍
B = tf.reshape(B_, [2, 4, 1])

# 或 更通用的改法：

#B_ = tf.tile(B, [tf.shape(A)[0], 1])

#B = tf.reshape(B_, [tf.shape(A)[0], tf.shape(B)[0], tf.shape(B)[1]])

#-----------------------------------------修改部分（结束）-----------------------------------------

#此时就可以matmul了
C = tf.matmul(A, B)
print(‘C:’,C.get_shape().as_list())
sess.run(C)

输出结果：

[python] view plain copy print?(‘C:’, [2, 3, 1])

array([[[30
aa9b
],
[70],
[20]],

[[20],
[60],
[40]]], dtype=int32)

('C:', [2, 3, 1])

array([[[30],
[70],
[20]],

[[20],
[60],
[40]]], dtype=int32)

转自博客：http://blog.csdn.net/blythe0107/article/details/74171870