Pytorch 学习（6）：Pytorch中的torch.nn Convolution Layers 卷积层参数初始化

版权声明：王家林大咖2018年新书《SPARK大数据商业实战三部曲》清华大学出版，清华大学出版社官方旗舰店（天猫）https://qhdx.tmall.com/?spm=a220o.1000855.1997427721.d4918089.4b2a2e5dT6bUsM https://blog.csdn.net/duan_zhihua/article/details/82596292

Pytorch 学习（6）：Pytorch中的torch.nn  Convolution Layers  卷积层参数初始化

class Conv1d(_ConvNd):
	......
    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
                 padding=0, dilation=1, groups=1, bias=True):
        kernel_size = _single(kernel_size)
        stride = _single(stride)
        padding = _single(padding)
        dilation = _single(dilation)
        super(Conv1d, self).__init__(
            in_channels, out_channels, kernel_size, stride, padding, dilation,
            False, _single(0), groups, bias)

    def forward(self, input):
        return F.conv1d(input, self.weight, self.bias, self.stride,
                        self.padding, self.dilation, self.groups)

参数初始化调用 _ntuple方法：

import collections
from itertools import repeat


def _ntuple(n):
    def parse(x):
        if isinstance(x, collections.Iterable):
            return x
        return tuple(repeat(x, n))
    return parse

_single = _ntuple(1)
_pair = _ntuple(2)
_triple = _ntuple(3)
_quadruple = _ntuple(4)

_ntuple是函数式编程高阶函数，_single = _ntuple(1)将n=1参数传入parse函数，返回parse函数，然后在_single(kernel_size)传入kernel_size参数，调用parse（kernel_size）方法，执行repeat(x, n)方法。

做个小测试：

import collections
from itertools import repeat
def _ntuple(n):
    def parse(x):
        if isinstance(x, collections.Iterable):
            return x
        return tuple(repeat(x, n))
    return parse

_single = _ntuple(1)
print(_single(0))

_pair = _ntuple(2)
print(_pair(0))
#kernel_size=5
kernel_size=(3, 5)
kernel_size = _pair(kernel_size)
print(kernel_size)

_triple = _ntuple(3)
kernel_size=(3, 5, 2)
kernel_size = _triple(kernel_size)
print(kernel_size)

运行结果如下：

(0,)
(0, 0)
(3, 5)
(3, 5, 2)