Learning letter cup -train code learning

F.nll_loss What is it?

negative likelihood loss
2. nll_loss , The full name of the function is negative log likelihood loss, The function expression is

f(x,class)=-x[class]

for example ： hypothesis x=[1,2,3], class=2, be

f(x,class)=-x[2]=-3

pytorch in .item() The role of

pytorch in ,.item() Method Is to get the element value in an element tensor
The concrete is Used to convert a zero dimensional tensor into a floating point number , For example, calculation loss,accuracy Value
Like ：
loss = (y_pred - y).pow(2).sum()
print(loss.item())

@torch.no_grad() and model.eval() The difference between

Be wrapped Forward calculation will be carried out for the part of , But it doesn't involve gradient calculation , Speed up calculation , Save computing resources .
because def test Only forward inference, Back propagation is not involved , There's no need to calculate the gradient , Therefore, using torch.no_grad()

model.eval()
The main function is to ensure dropout and batchnorm The normal operation ;
And ensure no back propagation ;
Although there is no back propagation ; However, this operation cannot prevent the calculation of gradient in forward propagation ; use @torch.no_grad() It can prevent the waste of resources caused by gradient calculation in forward calculation
https://blog.csdn.net/weixin_46559271/article/details/10565865

parser = argparse.ArgumentParser Use

model.state_dict() effect

model.reset_parameters()

def test

there test Function is very interesting , Not only forget it test, Also put train、valid and test It's all calculated , Easy to print .