对pytorch网络层结构的数组化详解

最近再写openpose，它的网络结构是多阶段的网络，所以写网络的时候很想用列表的方式，但是直接使用列表不能将网络中相应的部分放入到cuda中去。

其实这个问题很简单的，使用modulelist就好了。

1 我先是定义了一个函数，用来根据超参数，建立一个基础网络结构

stage = [[3, 3, 3, 1, 1], [7, 7, 7, 7, 7, 1, 1]]
branches_cfg = [[[128, 128, 128, 512, 38], [128, 128, 128, 512, 19]],
    [[128, 128, 128, 128, 128, 128, 38], [128, 128, 128, 128, 128, 128, 19]]]

# used for add two branches as well as adapt to certain stage
def add_extra(i, branches_cfg, stage):
 """
 only add cnn of brancdes s & l in stage ti at the end of net
 :param in_channels:the input channels & out
 :param stage: size of filter
 :param branches_cfg: channels of image
 :return:list of layers
 """
 in_channels = i
 layers = []
 for k in range(len(stage)):
  padding = stage[k] // 2
  conv2d = nn.conv2d(in_channels, branches_cfg[k], kernel_size=stage[k], padding=padding)
  layers += [conv2d, nn.relu(inplace=true)]
  in_channels = branches_cfg[k]
 return layers

2 然后用普通列表装载他们

conf_bra_list = []
paf_bra_list = []

# param for branch network
in_channels = 128

for i in range(all_stage):
 if i > 0:
  branches = branches_cfg[1]
  conv_sz = stage[1]
 else:
  branches = branches_cfg[0]
  conv_sz = stage[0]

 conf_bra_list.append(nn.sequential(*add_extra(in_channels, branches[0], conv_sz)))
 paf_bra_list.append(nn.sequential(*add_extra(in_channels, branches[1], conv_sz)))
 in_channels = 185

3 再然后，使用modulelist方法，把普通列表专成pytorch下的模块

# to list
self.conf_bra = nn.modulelist(conf_bra_list)
self.paf_bra = nn.modulelist(paf_bra_list)

4 最后，调用就好了

out_0 = x
# the base transform
for k in range(len(self.vgg)):
 out_0 = self.vgg[k](out_0)

# local name space
name = locals()
confs = []
pafs = []
outs = []

length = len(self.conf_bra)
for i in range(length):
 name['conf_%s' % (i + 1)] = self.conf_bra[i](name['out_%s' % i])
 name['paf_%s' % (i + 1)] = self.paf_bra[i](name['out_%s' % i])
 name['out_%s' % (i + 1)] = torch.cat([name['conf_%s' % (i + 1)], name['paf_%s' % (i + 1)], out_0], 1)
 confs.append('conf_%s' % (i + 1))
 pafs.append('paf_%s' % (i + 1))
 outs.append('out_%s' % (i + 1))

5 顺便装了一下，使用了python局部变量命名空间，name = locals()，其实完全使用普通列表保存变量就好了，高兴就好。

以上这篇对pytorch网络层结构的数组化详解就是小编分享给大家的全部内容了，希望能给大家一个参考，也希望大家多多支持。