发现自己干的事情和兔叭咯似的，沉迷挖坑
函数总结- torch.nn.Identity(): 返回原输入，用于保持网络层数
- torch.nn.MoudleList(): 构建一个存放layer，module的list，有利于forward的简写 -append和extend可以添加model进入list -e.g.
class model2(nn.Module):
def __init__(self):
super(model2, self).__init__()
self.layers=nn.ModuleList([
nn.Linear(1,10), nn.ReLU(),
nn.Linear(10,100),nn.ReLU(),
nn.Linear(100,10),nn.ReLU(),
nn.Linear(10,1)])
def forward(self,x):
out=x
for i,layer in enumerate(self.layers):
out=layer(out)
return out
self.linears.extend([nn.Linear(size1, size2) for i in range(1, num_layers)])
self.linears.append(nn.Linear(size1, size2)
- torch.nn.Sequential()：与modulelist类似，添加之后可以是做一个整体 -e.g.
class model3(nn.Module):
def __init__(self):
super(model3, self).__init__()
self.network=nn.Sequential(
nn.Linear(1,10),nn.ReLU(),
nn.Linear(10,100),nn.ReLU(),
nn.Linear(100,10),nn.ReLU(),
nn.Linear(10,1)
)
def forward(self, x):
return self.network(x)
#可以在创建之后加入新的模型，而且可以给子模型定义唯一的名称索引，方便获取
self.network.add_module("linear1",nn.Linear(100,100))
#获得这个子模型
linear=self.network.linear1
#另一种写法
from collections import OrderedDict
self.network=nn.Sequential(OrderedDict(
("linear1",nn.Linear(1,10)),
("activation1",nn.ReLU()),
("linear2",nn.Linear(10,100))
))
- model.to(device): 将模型加载到指定设备上
- optimizer.zero_grad()：梯度清零， 通过grad accumulate可以节约显存
for i,(image, label) in enumerate(train_loader):
# 1. input output
pred = model(image)
loss = criterion(pred, label)
# 2.1 loss regularization
loss = loss / accumulation_steps
# 2.2 back propagation
loss.backward()
# 3. update parameters of net
if (i+1) % accumulation_steps == 0:
# optimizer the net
optimizer.step() # update parameters of net
optimizer.zero_grad() # reset gradient
- loss = criterion(preds, targets) ：求解loss
- loss.backward() ：反向传播求解梯度
- optimizer.step()：更新模型参数
- with torch.no_grad(): warp内的计算过程不会被torch track
