CODEWITHSUNDEEP
pythonnumpy
Karnivaurus
Karnivaurus

Reputation: 24151

NumPy error saying that matrices are not aligned

If I run:

x = np.zeros(6)
y = np.zeros([7, 6])
z = y * x

Then everything is fine, and there are no Python errors.

However, I am using a Python module (call if foo) containing a function (call if bar), which returns a 7x6 NumPy array. It has the same shape as y above, and the same data type (float64). But when I run the following:

x = np.zeros(6)
y = foo.bar()
z = y * x

I get the following error:

ValueError: shapes (7,6) and (1,6) not aligned: 6 (dim 1) != 1 (dim 0)

But as far as I can tell, y is exactly the same format in these two examples, with the same shape and data type. What's causing this error, and why is it not caused in the first example?

Upvotes: 0

Views: 529

Answers (2)

hpaulj
hpaulj

Reputation: 231665

In [446]: x = np.zeros(6) 
     ...: y = np.zeros([7, 6]) 
     ...: z = y * x                                                                                    
In [447]: z.shape                                                                                      
Out[447]: (7, 6)

Here we are doing element-wise multiplication, a (7,6) with a (6,). By broadcasting the (6,) becomes (1,6) and then (7,6) to match y.

Evidently in the foo.bar case, y is np.matrix subclass:

In [454]: y1 = np.matrix(y)                                                                            
In [455]: y1*x                                                                                         
---...
    219             # This promotes 1-D vectors to row vectors
--> 220             return N.dot(self, asmatrix(other))
...
ValueError: shapes (7,6) and (1,6) not aligned: 6 (dim 1) != 1 (dim 0)

Note the different display for y1:

In [456]: y1                                                                                           
Out[456]: 
matrix([[0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.]])

With np.matrix * is defined as np.dot, the matrix product. The x is also converted np.matrix, producing a (1,6) matrix. The error message follows from the definition of matrix multiplication.

np.multiply can be used force the element-wise multiplication. Note the class of the result:

In [458]: np.multiply(y1,x)                                                                            
Out[458]: 
matrix([[0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.]])

Because of confusions like this np.matrix is being discouraged.

Upvotes: 1

Eric Grinstein
Eric Grinstein

Reputation: 339

I don't know which version you're running, but I am running version 1.16.3 on Python 3.

It seems to me you're defining your x differently than in your example snippet. You seem to be defining it as a 6x1 matrix instead of a "vector", which is considered on Numpy to only have one dimension. Try multiplying y by np.zeros([6,1]) and you'll see an error.

Bottom line is:

  1. Use the .shape property a lot when debugging, it's very useful when you're doing matrix multiplication.
  2. On np, multiplication between a matrix and a one dimensional array is carried differently than multiplications between 2 matrices.

Upvotes: 1

Related Questions