PyTorch 深度学习实践 第5讲 刘二大人 课后题个人答案

PyTorch 深度学习实践 第5讲 课后习题

由于本节习题很少看到比较完整的代码, 因此上传一个给大家参考.
B站 刘二大人

这里要注意一下, 在对不同优化器进行比较的过程中, 需要保持初始参数值一致, 这样才有可比性

# 练习5-1 尝试不同优化器训练线性模型,并保存图像
"""
Adagrad
Adam
Adamax"
ASGD
LBFGS
RMSprop
Rprop
SGD
"""

import torch
import matplotlib.pyplot as plt

x_data = torch.Tensor([[1.0], [2.0], [3.0]])
y_data = torch.Tensor([[2.0], [4.0], [6.0]])


class LinearModel(torch.nn.Module):
    def __init__(self, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)
        self.linear = torch.nn.Linear(1, 1)
        self.linear.weight.requires_grad = True
        self.linear.bias.requires_grad = True

    def forward(self, x):
        x = self.linear(x)
        return x


def init_weights(m):
    m.linear.weight.data.fill_(1.0)
    m.linear.bias.data.fill_(0.0)


model = LinearModel()

criterion = torch.nn.MSELoss(reduction='sum')


def optim_select(name: str):
    name_list = [
        "Adagrad", "Adam", "Adamax", "ASGD", "LBFGS", "RMSprop", "Rprop", "SGD"
    ]
    if name == "LBFGS":
        return eval(f"torch.optim.{name}(model.parameters(), lr=0.001)")
    if name in name_list:
        return eval(f"torch.optim.{name}(model.parameters(), lr=0.01)")
    else:
        raise NameError(f"{name}不是一个支持的优化器")


plt.figure(num='fig1')
for optime in ["Adagrad", "Adam"]:
    print(f"当前优化器为{optime}")
    init_weights(model)
    optimezer = optim_select(optime)
    l_list = []
    for epoch in range(1000):
        y_pred = model(x_data)
        l = criterion(y_pred, y_data)
        l_list.append(l.item())
        if epoch % 200 == 0:
            print(f"{epoch} -- {l.item():.3f}")
        optimezer.zero_grad()
        l.backward()
        optimezer.step()

    x_test = torch.Tensor([[4.0]])
    y_test = model(x_test)
    print("y_pred =", y_test.data.item())

    plt.plot(range(epoch + 1), l_list, label=f"{optime}")
    plt.xlabel("epoch")
    plt.ylabel("loss")
    plt.title("epoch-loss")
    plt.legend()
plt.show()

plt.figure(num='fig2')
for optime in ["Adamax", "ASGD", "RMSprop", "Rprop", "SGD"]:
    print(f"当前优化器为{optime}")
    init_weights(model)
    optimezer = optim_select(optime)
    l_list = []
    for epoch in range(100):
        y_pred = model(x_data)
        l = criterion(y_pred, y_data)
        l_list.append(l.item())
        if epoch % 20 == 0:
            print(f"{epoch} -- {l.item():.3f}")
        optimezer.zero_grad()
        l.backward()
        optimezer.step()
    x_test = torch.Tensor([[4.0]])
    y_test = model(x_test)
    print("y_pred =", y_test.data.item())

    plt.plot(range(epoch + 1), l_list, label=f"{optime}")
    plt.xlabel("epoch")
    plt.ylabel("loss")
    plt.title("epoch-loss")
    plt.legend()
plt.show()


def closure():  # - LBFGS需要传入闭包函数作为参数
    y_pred = model(x_data)
    loss = criterion(y_pred, y_data)
    return loss


l_list = []
plt.figure(num='fig3')
for optime in ["LBFGS"]:
    print(f"当前优化器为{optime}")
    init_weights(model)
    optimezer = optim_select(optime)
    for epoch in range(1000):
        optimezer.zero_grad()
        loss = closure()
        if epoch % 100 == 0:
            print(f"{epoch} -- {loss.item():.3f}")
        loss.backward()
        l_list.append(loss.item())
        optimezer.step(closure)

    x_test = torch.Tensor([[4.0]])
    y_test = model(x_test)
    print("y_pred =", y_test.data.item())

    plt.plot(range(epoch + 1), l_list, label=f"{optime}")
    plt.xlabel("epoch")
    plt.ylabel("loss")
    plt.title("epoch-loss")
    plt.legend()
    plt.text(600, 8, f'traing epoch: {epoch+1}\nlearning rate: 0.001')

plt.show()

图片不知为啥上传不了, 大家可以copy代码本地运行查看结果