pytorch resnet實(shí)現(xiàn)
官方github上已經(jīng)有了pytorch基礎(chǔ)模型的實(shí)現(xiàn)，鏈接
但是其中一些模型，尤其是resnet，都是用函數(shù)生成的各個(gè)層，自己看起來是真的難受！
所以自己按照caffe的樣子，寫一個(gè)pytorch的resnet18模型，當(dāng)然和1000分類模型不同，模型做了一些修改，輸入48*48的3通道圖片，輸出7類。
import torch.nn as nn import torch.nn.functional as F class ResNet18Model(nn.Module): def __init__(self): super().__init__() self.bn64_0 = nn.BatchNorm2d(64) self.bn64_1 = nn.BatchNorm2d(64) self.bn64_2 = nn.BatchNorm2d(64) self.bn64_3 = nn.BatchNorm2d(64) self.bn64_4 = nn.BatchNorm2d(64) self.bn128_0 = nn.BatchNorm2d(128) self.bn128_1 = nn.BatchNorm2d(128) self.bn128_2 = nn.BatchNorm2d(128) self.bn128_3 = nn.BatchNorm2d(128) self.bn256_0 = nn.BatchNorm2d(256) self.bn256_1 = nn.BatchNorm2d(256) self.bn256_2 = nn.BatchNorm2d(256) self.bn256_3 = nn.BatchNorm2d(256) self.bn512_0 = nn.BatchNorm2d(512) self.bn512_1 = nn.BatchNorm2d(512) self.bn512_2 = nn.BatchNorm2d(512) self.bn512_3 = nn.BatchNorm2d(512) self.shortcut_straight_0 = nn.Sequential() self.shortcut_straight_1 = nn.Sequential() self.shortcut_straight_2 = nn.Sequential() self.shortcut_straight_3 = nn.Sequential() self.shortcut_straight_4 = nn.Sequential() self.shortcut_conv_bn_64_128_0 = nn.Sequential(nn.Conv2d(64, 128, kernel_size=1, stride=2, bias=False),nn.BatchNorm2d(128)) self.shortcut_conv_bn_128_256_0 = nn.Sequential(nn.Conv2d(128, 256, kernel_size=1, stride=2, bias=False),nn.BatchNorm2d(256)) self.shortcut_conv_bn_256_512_0 = nn.Sequential(nn.Conv2d(256, 512, kernel_size=1, stride=2, bias=False),nn.BatchNorm2d(512)) self.conv_w3_h3_in3_out64_s1_p1_0 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in64_out64_s1_p1_0 = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in64_out64_s1_p1_1 = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in64_out64_s1_p1_2 = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in64_out64_s1_p1_3 = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in64_out128_s2_p1_0 = nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1, bias=False) self.conv_w3_h3_in128_out128_s1_p1_0 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in128_out128_s1_p1_1 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in128_out128_s1_p1_2 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in128_out256_s2_p1_0 = nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1, bias=False) self.conv_w3_h3_in256_out256_s1_p1_0 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in256_out256_s1_p1_1 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in256_out256_s1_p1_2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in256_out512_s2_p1_0 = nn.Conv2d(256, 512, kernel_size=3, stride=2, padding=1, bias=False) self.conv_w3_h3_in512_out512_s1_p1_0 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in512_out512_s1_p1_1 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1, bias=False) self.conv_w3_h3_in512_out512_s1_p1_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1, bias=False) self.avg_pool_0 = nn.AdaptiveAvgPool2d((1, 1)) self.fc_512_7_0 = nn.Linear(512, 7) self.dropout_0 = nn.Dropout(p=0.5) def forward(self, x): # 48*48*3 t = self.conv_w3_h3_in3_out64_s1_p1_0(x) #48*48*64 t = self.bn64_0(t) y1 = F.relu(t) t = self.conv_w3_h3_in64_out64_s1_p1_0(y1) #48*48*64 t = self.bn64_1(t) y2 = F.relu(t) t = self.conv_w3_h3_in64_out64_s1_p1_1(y2) #48*48*64 t = self.bn64_2(t) t += self.shortcut_straight_0(y1) y3 = F.relu(t) t = self.conv_w3_h3_in64_out64_s1_p1_2(y3) #48*48*64 t = self.bn64_3(t) y4 = F.relu(t) t = self.conv_w3_h3_in64_out64_s1_p1_3(y4) #48*48*64 t = self.bn64_4(t) t += self.shortcut_straight_1(y3) y5 = F.relu(t) t = self.conv_w3_h3_in64_out128_s2_p1_0(y5) #24*24*128 t = self.bn128_0(t) y6 = F.relu(t) t = self.conv_w3_h3_in128_out128_s1_p1_0(y6) #24*24*128 t = self.bn128_1(t) t += self.shortcut_conv_bn_64_128_0(y5) y7 = F.relu(t) t = self.conv_w3_h3_in128_out128_s1_p1_1(y7) #24*24*128 t = self.bn128_2(t) y8 = F.relu(t) t = self.conv_w3_h3_in128_out128_s1_p1_2(y8) #24*24*128 t = self.bn128_3(t) t += self.shortcut_straight_2(y7) y9 = F.relu(t) t = self.conv_w3_h3_in128_out256_s2_p1_0(y9) #12*12*256 t = self.bn256_0(t) y10 = F.relu(t) t = self.conv_w3_h3_in256_out256_s1_p1_0(y10) #12*12*256 t = self.bn256_1(t) t += self.shortcut_conv_bn_128_256_0(y9) y11 = F.relu(t) t = self.conv_w3_h3_in256_out256_s1_p1_1(y11) #12*12*256 t = self.bn256_2(t) y12 = F.relu(t) t = self.conv_w3_h3_in256_out256_s1_p1_2(y12) #12*12*256 t = self.bn256_3(t) t += self.shortcut_straight_3(y11) y13 = F.relu(t) t = self.conv_w3_h3_in256_out512_s2_p1_0(y13) #6*6*512 t = self.bn512_0(t) y14 = F.relu(t) t = self.conv_w3_h3_in512_out512_s1_p1_0(y14) #6*6*512 t = self.bn512_1(t) t += self.shortcut_conv_bn_256_512_0(y13) y15 = F.relu(t) t = self.conv_w3_h3_in512_out512_s1_p1_1(y15) #6*6*512 t = self.bn512_2(t) y16 = F.relu(t) t = self.conv_w3_h3_in512_out512_s1_p1_2(y16) #6*6*512 t = self.bn512_3(t) t += self.shortcut_straight_4(y15) y17 = F.relu(t) out = self.avg_pool_0(y17) #1*1*512 out = out.view(out.size(0), -1) out = self.dropout_0(out) out = self.fc_512_7_0(out) return out if __name__ == '__main__': net = ResNet18Model() # print(net) import torch net_in = torch.rand(1, 3, 48, 48) net_out = net(net_in) print(net_out) print(net_out.size())
浙公網(wǎng)安備 33010602011771號