Break training data into slices for stochastic gradient decent:

import numpy as np
n = 100
training_data = list(range(n))
mini_batch_size = 10
np.random.shuffle(training_data)
mini_batches = [training_data[k:k+mini_batch_size]
    for k in range(0, n, mini_batch_size)]
mini_batches

[[90, 5, 70, 82, 58, 2, 16, 85, 12, 35],
 [14, 54, 62, 39, 96, 73, 60, 80, 33, 89],
 [20, 38, 76, 47, 65, 42, 71, 46, 93, 34],
 [52, 64, 13, 92, 17, 49, 88, 63, 74, 23],
 [43, 25, 10, 97, 48, 68, 95, 81, 24, 31],
 [9, 32, 84, 83, 22, 87, 61, 26, 28, 99],
 [0, 67, 30, 69, 72, 45, 79, 51, 40, 55],
 [6, 15, 75, 66, 29, 3, 18, 77, 98, 21],
 [53, 44, 50, 19, 91, 8, 11, 59, 27, 56],
 [36, 94, 7, 57, 1, 37, 86, 78, 41, 4]]

numpy.array vs numpy.asarray

Looking at the definition, you'll see the difference between them:

def asarray(a, dtype=None, order=None):
    return array(a, dtype, copy=False, order=order)

The main difference is that array (by default) will make a copy of the object, while asarray will not unless necessary.

The difference can be demonstrated by this example:

1. generate a matrix

   >>> A = numpy.matrix(np.ones((3,3)))
   >>> A
   matrix([[ 1.,  1.,  1.],
           [ 1.,  1.,  1.],
           [ 1.,  1.,  1.]])

2. use numpy.array to modify A. Doesn't work because you are modifying a copy

   >>> numpy.array(A)[2]=2
   >>> A
   matrix([[ 1.,  1.,  1.],
           [ 1.,  1.,  1.],
           [ 1.,  1.,  1.]])

3. use numpy.asarray to modify A. It worked because you are modifying A itself

   >>> numpy.asarray(A)[2]=2
   >>> A
   matrix([[ 1.,  1.,  1.],
           [ 1.,  1.,  1.],
           [ 2.,  2.,  2.]])