当前位置:网站首页>Mixconv code
Mixconv code
2022-07-01 22:42:00 【zouxiaolv】
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
def split_layer(total_channels, num_groups):
# print('total_channels', 'num_groups',total_channels, num_groups)
a =[]
for _ in range(num_groups):
a.append(int(np.ceil(total_channels / num_groups)))
# print('a=',a)
split = [int(np.ceil(total_channels / num_groups)) for _ in range(num_groups)]
# print('split= ',split)
# print('splitsum= ', total_channels - sum(split))
split[num_groups - 1] += total_channels - sum(split)
# print('split111= ', split)
return split
class DepthwiseConv2D(nn.Module):
def __init__(self, in_channels, kernal_size, stride, bias=False):
super(DepthwiseConv2D, self).__init__()
padding = (kernal_size - 1) // 2
self.depthwise_conv = nn.Conv2d(in_channels, in_channels, kernel_size=kernal_size, padding=padding, stride=stride, groups=in_channels, bias=bias)
def forward(self, x):
out = self.depthwise_conv(x)
return out
class GroupConv2D(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=1, n_chunks=1, bias=False):
super(GroupConv2D, self).__init__()
self.n_chunks = n_chunks
self.split_in_channels = split_layer(in_channels, n_chunks)
# print('self.split_in_channels=',self.split_in_channels)
split_out_channels = split_layer(out_channels, n_chunks)
if n_chunks == 1:
self.group_conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias)
else:
self.group_layers = nn.ModuleList()
for idx in range(n_chunks):
self.group_layers.append(nn.Conv2d(self.split_in_channels[idx], split_out_channels[idx], kernel_size=kernel_size, bias=bias))
def forward(self, x):
if self.n_chunks == 1:
return self.group_conv(x)
else:
split = torch.split(x, self.split_in_channels, dim=1)
out = torch.cat([layer(s) for layer, s in zip(self.group_layers, split)], dim=1)
return out
class MDConv(nn.Module):
def __init__(self, out_channels, n_chunks, stride=1, bias=False):
super(MDConv, self).__init__()
self.n_chunks = n_chunks
self.split_out_channels = split_layer(out_channels, n_chunks)
print('self.split_out_channels=',self.split_out_channels)
self.layers = nn.ModuleList()
for idx in range(self.n_chunks):
kernel_size = 2 * idx + 3
self.layers.append(DepthwiseConv2D(self.split_out_channels[idx], kernal_size=kernel_size, stride=stride, bias=bias))
def forward(self, x):
split = torch.split(x, self.split_out_channels, dim=1)
out = torch.cat([layer(s) for layer, s in zip(self.layers, split)], dim=1)
return out
temp = torch.randn((16, 3, 32, 32))
# group = GroupConv2D(3, 16, n_chunks=2)
# print(group(temp).size())
group = MDConv(3, n_chunks=2)
print(group(temp).size())
Explanation is : First group the input , After subletting , Each group uses different convolution kernel sizes , Complete the convolution process with deep separable convolution
shortcoming : similar inception, The more groups , The faster the speed decreases . Only focus on parameter reduction
边栏推荐
猜你喜欢
随机推荐
比较版本号[双指针截取自己想要的字串]
100年仅6款产品获批,疫苗竞争背后的“佐剂”江湖
Appium automated testing foundation - Supplement: introduction to desired capabilities parameters
Matlab traverses images, string arrays and other basic operations
C#/VB.NET 给PDF文档添加文本/图像水印
隐藏用户的创建和使用
Mysql database detailed learning tutorial
分享一个一年经历两次裁员的程序员的一些感触
Fully annotated SSM framework construction
陈天奇的机器学习编译课(免费)
#yyds干货盘点# 解决名企真题:扭蛋机
Compensation des créneaux horaires
从零开始学 MySQL —数据库和数据表操作
Learning notes on futuretask source code of concurrent programming series
快乐数[环类问题之快慢指针]
删除AWS绑定的信用卡账户
MySQL中对于事务的理解
SAP UI5 应用开发教程之一百零四 - SAP UI5 表格控件的支持复选(Multi-Select)以及如何用代码一次选中多个表格行项目
Redis configuration and optimization
三翼鸟两周年:羽翼渐丰,腾飞指日可待