有效地从数组中获取批量元素 - Python

时间:2017-04-28 10:02:24

标签: python numpy vectorization

我是Python和Numpy的新手。

我有这个输入:

[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15]

我想得到这个输出:

[
  [
    [[ 1  2  3], [ 7  8  9]],
    [[ 2  3  4], [ 8  9 10]]
  ],

  [
    [[ 4  5  6], [10 11 12]],
    [[ 5  6  7], [11 12 13]]
  ]
]

该数组具有以下形状:(n_batches, 2, batch_size, seq_length)。在下面的代码中,您可以看到这些参数的含义。

我是使用for loops完成的。请注意,以下代码来自here

def get_batches(int_text, batch_size, seq_length):
    """
    Return batches of input and target
    :param int_text: Text with the words replaced by their ids
    :param batch_size: The size of batch
    :param seq_length: The length of sequence
    :return: Batches as a Numpy array
    """
    # TODO: Implement Function    
    n_batches = int(len(int_text) / batch_size / seq_length)    
    batches = np.zeros(shape = (n_batches, 2, batch_size, seq_length), dtype = np.int32)
    index = 0
    for batch_index in range(n_batches):  
        one_batch = np.zeros(shape = (2, batch_size, seq_length), dtype = np.int32)
        one_batch_seq = np.zeros(shape = (batch_size, seq_length), dtype = np.int32)
        one_batch_tar = np.zeros(shape = (batch_size, seq_length), dtype = np.int32)
        for n in range(batch_size):
            index = batch_index * seq_length + n * n_batches * seq_length            
            seq = np.array([int_text[index:index+seq_length]])
            tar = np.array([int_text[index+1:index+seq_length+1]])
            one_batch_seq[n] = seq;
            one_batch_tar[n] = tar;
            #print(seq)
            #if(one_batch_seq.size == 0):
            #    one_batch_seq = seq
            #else:
            #    one_batch_seq = np.concatenate((one_batch_seq, seq))
            #if(one_batch_tar.size == 0):
            #    one_batch_tar = tar
            #else:
            #    one_batch_tar = np.concatenate((one_batch_tar, tar))    
        one_batch[0] = one_batch_seq
        one_batch[1] = one_batch_tar
        batches[batch_index] = one_batch

    return batches

代码说明

我必须得到我可以使用该输入完成的批次数(int_text)。我得到值n_batches来了解这一点。如果无法使用足够的数据填充最后一批,请删除最后一批。

然后,我创建一个形状为(n_batches, 2, batch_size, seq_length)的数组。然后,循环。

但有人告诉我,我可以不用循环而只使用Numpy。我不知道该怎么做。我想,我可以用重塑来做,但我不确定。

如何在不使用for-loop的情况下获得该输出?

1 个答案:

答案 0 :(得分:1)

这是一个超级高效的NumPy strides -

def get_batches_strided(a, batch_size, seq_length):
    n_batches = int(len(a) / batch_size / seq_length)   
    shp = (n_batches, 2, batch_size, seq_length)

    n = a.strides[0]
    s = np.lib.stride_tricks.as_strided
    strides=(shp[3]*n,n,shp[0]*shp[3]*n,n)
    out = s(a, shp, strides=strides)
    return out

这将是输入数组的视图,因此不会占用更多的内存空间,而且它也是它提高效率的地方。

此外,由于我们正在使用views并且基于strides的方法将超出为输入数组分配的内存,如果形状/ strides如此列出,我们需要注意它的输入参数。

运行示例以验证各种输入参数 -

In [306]: a = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15])

# List of values to select batch_size, seq_length to form different combinations
# of inputs to be fed to the two proposed approaches for tesing
In [308]: r = [2,3,4]

In [309]: [np.allclose(get_batches(a,m,n), get_batches_strided(a,m,n)) \
                                                    for m in r for n in r]
Out[309]: [True, True, True, True, True, True, True, True, True] # All checked out

对性能改进的一瞥 -

In [320]: a = np.arange(1,2001)

In [321]: %timeit get_batches(a, 3,2)
100 loops, best of 3: 2.87 ms per loop

In [322]: %timeit get_batches_strided(a, 3,2)
100000 loops, best of 3: 5.07 µs per loop

如果我们需要副本,那就不会太昂贵 -

In [323]: %timeit get_batches_strided(a, 3,2).copy()
100000 loops, best of 3: 15.4 µs per loop

工作流程回顾决定strides

使用np.lib.stride_tricks.as_strided时,只有两个关键参数 - 形状和步幅。形状还可以,因为我们可以从循环版本中获得。取得进展是棘手的,并且有大量的反复试验决定它。

让我尝试播放它,以帮助尝试学习/使用strides的人。

输入是:

a = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15])

获取步幅并为步长func指定一个较短的名称:

n = a.strides[0]
s = np.lib.stride_tricks.as_strided

测试了各种步幅组合的跨步方法,下面列出了每个输入对的成功方法 -

#get_batches(a, 2, 2) # case1
#s(a, (3,2,2,2), strides=(2*n,1*n,6*n,n)) # worked

#get_batches(a, 2, 3) # case2
#s(a, (2,2,2,3), strides=(3*n,1*n,6*n,n)) # worked

#get_batches(a, 3,4) # case3
#s(a, (1,2,3,4), strides=(4*n,1*n,4*n,n)) # worked

#get_batches(a, 3,2) # case4
#s(a, (2,2,3,2), strides=(2*n,1*n,4*n,n)) # worked

接下来的工作是找到这种模式,经过一些看起来很明显:

#strides=(shp[3]*n,n,shp[0]*shp[3]*n,n), where shp is the input shape tuple