如何展平（n，）形状numpy数组

Question

I have a variable called data, and i want to flatten it.

Right now data.shape = (10, ) . Each element in data has a shape (5000, 64). I want to make my data.shape = (10, 5000, 64) .

How can I do that? I've tried many below, but none of them work:

b = np.concatenate(t for t in data)
b = np.stack(t for t in data)
b = np.hstack(t for t in data)

anyone any idea about this?

Answer 1

In [50]: t = np.zeros(3,object)
In [51]: t[0]=np.ones((3,2),int)
In [52]: t[1]=np.ones((3,2),int)
In [53]: t[2]=np.ones((3,2),int)
In [54]: t
Out[54]: 
array([array([[1, 1],
       [1, 1],
       [1, 1]]),
       array([[1, 1],
       [1, 1],
       [1, 1]]),
       array([[1, 1],
       [1, 1],
       [1, 1]])], dtype=object)
In [55]: np.stack(t)
Out[55]: 
array([[[1, 1],
        [1, 1],
        [1, 1]],

       [[1, 1],
        [1, 1],
        [1, 1]],

       [[1, 1],
        [1, 1],
        [1, 1]]])
In [56]: _.shape
Out[56]: (3, 3, 2)

np.stack(x for x in t) also works, but the extra layer of iteration isn't needed.

When you say something isn't working, you need to elaborate. What exactly was the error? Was it the same error for each of things you tried? Without information like that you'll get answers like mine - a test case that appears to match your description which works.

hstack works, but joins the arrays in a different way.

concatenate doesn't work with the (t for t in x) input, but is ok with (x) .

Assuming my guess about it being a object dtype array is right, this isn't a case of flatten . numpy flatten is like ravel, a shape changer. But this is different. I have arrays inside an object array, effectively a list of arrays, not a multidimensional array.

Answer 2

I solved this problem. The problem is that my data is a numpy array of tensors, and you have to use tf.stack to do the thing.