2

I'm trying to create np.array of np.arrays of different shape. I need it because in the next part I will sum up this large np.array with some delta_array that has the same shape. Like matrices sum self.weights += delta_weights And I will be working with each array inside it separately as well. Or for example I will need to multiply element wise all arrays by some number.

Can't figure out what's wrong :( Please, help me. Code creates list of random np.arrays of different shape. So here is my code:

weights = []
for i in range(1, len(self.layers)):
    weights.append(np.random.rand(self.layers[i-1] + 1, self.layers[i]))
print(type(weights))
print([(type(w), w.shape) for w in weights])
#error here with layers = [2,2,1] or [3,3,1] etc
self.weights = np.array(weights)

Output: For self.layers=[2, 2, 1]

<class 'list'>
[(<class 'numpy.ndarray'>, (3, 2)), (<class 'numpy.ndarray'>, (3, 1))]
Traceback (most recent call last):
line 20, in <module>
 run()
line 8, in run
 net.init_weights()
line 71, in init_weights
 self.weights = np.array(weights)
ValueError: could not broadcast input array from shape (3,2) into shape (3)

For [2, 3, 1] everything is okay:

<class 'list'>
[(<class 'numpy.ndarray'>, (3, 3)), (<class 'numpy.ndarray'>, (4, 1))]

For [3, 3, 1] same story as for [2, 2, 1] - error

For [3, 7, 1] or [3, 2, 1] everything is okay.

//It's matrices of weights for gradient descent for machine learning.

userqwerty1
  • 887
  • 2
  • 9
  • 23

2 Answers2

4

Ok, you're trying to create an array of matrices, so that you can use array operations (like +=) later on.

You can declare an ndarray to hold another ndarray, and then fill it with the matrices (which are simply two dimensional ndarrays):

weights = np.empty(len(self.layers), dtype=np.ndarray)
for i in range(1, len(self.layers)):
    weights[i] = np.random.rand(self.layers[i-1] + 1, self.layers[i])

Keep in mind that this is definitely not a three dimensional array: it's purely an array of two dimensional arrays.

The reason your code works for [2, 3, 1] is that the array with one dimension of 1 gets broadcasted; that's purely accidental, and not what you want, as far as I understand.

2

When you give np.array a list of arrays, it tries to combine them into one array (with a higher ndim). Only if it can't combine them will it create an array of dtype object with one array in each slot.

In the [2,3,1] case the 2 arrays are (3,3) and (4,1). Since there's no way to 'stack' these, it creates a 2 element array containing these 2.

In the [3,3,1] case the 2 arrays are (4,3) and (4,1). Due to an overlap in dimensions (4 rows) it appears to be trying to join them into one array, possibly (2,4,?), but isn't quite succeeding. In your version it raises an error, in my version it creates a (2,4) array of Nones.

Evert's solution is a general purpose one that correctly handles these confusing case. It works even if the 2 arrays in weights were the same size.

Say, for example if your weights contained 2 (4,3) arrays. np.arrays(weights) would produce a (2,4,3) array. Evert's solution would still produce a (2,) array containing 2 (4,3) arrays.

Looks like there's a bug (or two) in different numpy versions. Python3 with 1.8.0 numpy handled the [2,2,1] case just fine.

A simpler test case would be: 

np.array([np.ones((2,2)), np.zeros((2,1))])

Playing around the dimensions of the 2 arrays. It should produce a 2 element array in most cases, and a (2,n,m) array if the 2 dimensions are the same.

hpaulj
  • 221,503
  • 14
  • 230
  • 353