創(chuàng)新互聯(lián)www.cdcxhl.cn八線動態(tài)BGP香港云服務器提供商，新人活動買多久送多久，劃算不套路！

成都創(chuàng)新互聯(lián)公司-專業(yè)網(wǎng)站定制、快速模板網(wǎng)站建設、高性價比定興網(wǎng)站開發(fā)、企業(yè)建站全套包干低至880元,成熟完善的模板庫,直接使用。一站式定興網(wǎng)站制作公司更省心,省錢,快速模板網(wǎng)站建設找我們，業(yè)務覆蓋定興地區(qū)。費用合理售后完善，10余年實體公司更值得信賴。

這期內(nèi)容當中小編將會給大家?guī)碛嘘Ppytorch實現(xiàn)unet網(wǎng)絡的方法，文章內(nèi)容豐富且以專業(yè)的角度為大家分析和敘述，閱讀完這篇文章希望大家可以有所收獲。

設計神經(jīng)網(wǎng)絡的一般步驟：

1. 設計框架

2. 設計骨干網(wǎng)絡

Unet網(wǎng)絡設計的步驟：

1. 設計Unet網(wǎng)絡工廠模式

2. 設計編解碼結(jié)構

3. 設計卷積模塊

4. unet實例模塊

Unet網(wǎng)絡最重要的特征：

1. 編解碼結(jié)構。

2. 解碼結(jié)構，比FCN更加完善，采用連接方式。

3. 本質(zhì)是一個框架，編碼部分可以使用很多圖像分類網(wǎng)絡。

示例代碼：

import torch
import torch.nn as nn

class Unet(nn.Module):
 #初始化參數(shù)：Encoder,Decoder,bridge
 #bridge默認值為無，如果有參數(shù)傳入，則用該參數(shù)替換None
 def __init__(self,Encoder,Decoder,bridge = None):
  super(Unet,self).__init__()
  self.encoder = Encoder(encoder_blocks)
  self.decoder = Decoder(decoder_blocks)
  self.bridge = bridge
 def forward(self,x):
  res = self.encoder(x)
  out,skip = res[0],res[1,:]
  if bridge is not None:
   out = bridge(out)
  out = self.decoder(out,skip)
  return out
#設計編碼模塊
class Encoder(nn.Module):
 def __init__(self,blocks):
  super(Encoder,self).__init__()
  #assert:斷言函數(shù)，避免出現(xiàn)參數(shù)錯誤
  assert len(blocks) > 0
  #nn.Modulelist（）：模型列表，所有的參數(shù)可以納入網(wǎng)絡，但是沒有forward函數(shù)
  self.blocks = nn.Modulelist(blocks)
 def forward(self,x):
  skip = []
  for i in range(len(self.blocks) - 1):
   x = self.blocks[i](x)
   skip.append(x)
  res = [self.block[i+1](x)]
  #列表之間可以通過+號拼接
  res += skip
  return res
#設計Decoder模塊
class Decoder(nn.Module):
 def __init__(self,blocks):
  super(Decoder, self).__init__()
  assert len(blocks) > 0
  self.blocks = nn.Modulelist(blocks)
 def ceter_crop(self,skips,x):
  _,_,height1,width2 = skips.shape()
  _,_,height2,width3 = x.shape()
  #對圖像進行剪切處理，拼接的時候保持對應size參數(shù)一致
  ht,wt = min(height1,height2),min(width2,width3)
  dh2 = (height1 - height2)//2 if height1 > height2 else 0
  dw1 = (width2 - width3)//2 if width2 > width3 else 0
  dh3 = (height2 - height1)//2 if height2 > height1 else 0
  dw2 = (width3 - width2)//2 if width3 > width2 else 0
  return skips[:,:,dh2:(dh2 + ht),dw1:(dw1 + wt)],\
    x[:,:,dh3:(dh3 + ht),dw2 : (dw2 + wt)]

 def forward(self, skips,x,reverse_skips = True):
  assert len(skips) == len(blocks) - 1
  if reverse_skips is True:
   skips = skips[: : -1]
  x = self.blocks[0](x)
  for i in range(1, len(self.blocks)):
   skip = skips[i-1]
   x = torch.cat(skip,x,1)
   x = self.blocks[i](x)
  return x
#定義了一個卷積block
def unet_convs(in_channels,out_channels,padding = 0):
 #nn.Sequential：與Modulelist相比，包含了forward函數(shù)
 return nn.Sequential(
  nn.Conv2d(in_channels, out_channels, kernal_size = 3, padding = padding, bias = False),
  nn.BatchNorm2d(outchannels),
  nn.ReLU(inplace = True),
  nn.Conv2d(in_channels, out_channels, kernal_size=3, padding=padding, bias=False),
  nn.BatchNorm2d(outchannels),
  nn.ReLU(inplace=True),
 )
#實例化Unet模型
def unet(in_channels,out_channels):
 encoder_blocks = [unet_convs(in_channels, 64),\
      nn.Sequential(nn.Maxpool2d(kernal_size = 2, stride = 2, ceil_mode = True),\
         unet_convs(64,128)), \
      nn.Sequential(nn.Maxpool2d(kernal_size=2, stride=2, ceil_mode=True), \
         unet_convs(128, 256)),
      nn.Sequential(nn.Maxpool2d(kernal_size=2, stride=2, ceil_mode=True), \
         unet_convs(256, 512)),
      ]
 bridge = nn.Sequential(unet_convs(512, 1024))
 decoder_blocks = [nn.conTranpose2d(1024, 512), \
      nn.Sequential(unet_convs(1024, 512),
         nn.conTranpose2d(512, 256)),\
      nn.Sequential(unet_convs(512, 256),
         nn.conTranpose2d(256, 128)), \
      nn.Sequential(unet_convs(512, 256),
         nn.conTranpose2d(256, 128)), \
      nn.Sequential(unet_convs(256, 128),
         nn.conTranpose2d(128, 64))
      ]
 return Unet(encoder_blocks,decoder_blocks,bridge)

新聞標題：pytorch實現(xiàn)unet網(wǎng)絡的方法-創(chuàng)新互聯(lián)
本文URL：http://weahome.cn/article/pepjj.html