pytorch实践（更改resnet50全连接层，重新训练，用于二分类）

背景交代（很杂乱，可跳过）：

最开始的目的是完成和SAN论文里一样，用把四种风格图片当做四种标签，在原来的权重基础上，整体重新训练resnet50两轮（原文中是resnet101），下图为resnet网络结构

改变了最后的全连接层，从头开始重新训练，结果不太对，训练日志部分如下，一直不是等于100就是0

Training dir : D:\postgraduate\major_related\face_related\datasets\SAN-300W-test\300W-Convert
epoch : [0/50] lr=0.01
after 1 epoch, prec:57.14285714285714%,loss:1.261704683303833,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 1 epoch, prec:100.0%,loss:0.0,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 1 epoch, prec:0.0%,loss:23.00307846069336,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 1 epoch, prec:0.0%,loss:15.251273155212402,input:torch.Size([6, 3, 224, 224]),output:torch.Size([6, 4])
epoch : [1/50] lr=0.009978950082958632
after 2 epoch, prec:0.0%,loss:32.996482849121094,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 2 epoch, prec:0.0%,loss:20.09764289855957,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 2 epoch, prec:0.0%,loss:1.9084116220474243,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 2 epoch, prec:0.0%,loss:2.1563243865966797,input:torch.Size([6, 3, 224, 224]),output:torch.Size([6, 4])
epoch : [2/50] lr=0.00995794447581801
after 3 epoch, prec:0.0%,loss:6.454265594482422,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 3 epoch, prec:0.0%,loss:2.507550001144409,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 3 epoch, prec:0.0%,loss:2.3660380840301514,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 3 epoch, prec:0.0%,loss:2.6066017150878906,input:torch.Size([6, 3, 224, 224]),output:torch.Size([6, 4])
epoch : [3/50] lr=0.009936983085306158
after 4 epoch, prec:0.0%,loss:1.5964949131011963,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 4 epoch, prec:0.0%,loss:1.3409065008163452,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 4 epoch, prec:0.0%,loss:2.040916681289673,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 4 epoch, prec:0.0%,loss:1.7939444780349731,input:torch.Size([6, 3, 224, 224]),output:torch.Size([6, 4])
epoch : [4/50] lr=0.00991606581834744
after 5 epoch, prec:14.285714285714285%,loss:1.5786248445510864,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 5 epoch, prec:0.0%,loss:2.005967855453491,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 5 epoch, prec:0.0%,loss:1.800467848777771,input:torch.Size([7, 3, 224, 224]),output:torch.Size([7, 4])
after 5 epoch, prec:0.0%,loss:1.7157955169677734,input:torch.Size([6, 3, 224, 224]),output:torch.