Pytorch入门——基础知识及实现两层网络

Pytorch基础知识

内容来源：
B站视频——最好的PyTorch的入门与实战教程（16小时实战）

import torch
import numpy as np

torch.empty(5,3)  # 创建未初始化的矩阵

x1 = torch.rand(5,3)  # 随机初始化矩阵

x2 = torch.zeros(5,3)  # 全部为0矩阵

x3 = torch.zeros(5,3, dtype=torch.long)  # 数据类型变为long
# x3 = torch.zeros(5,3).long() 效果一样

x4 = torch.tensor([5.5, 3])  # 从数据直接构建tensor

x5 = x4.new_ones(5,3)  # 根据已有tensor构建一个tensor，这些方法会重用原来tensor的特征。例如数据类型

x6 = x4.new_ones(5,3, dtype=torch.double)

torch.rand_like(x5, dtype=torch.float)

# 得到tensor的形状
x5.shape
x5.size

# 运算
y1 = torch.rand(5,3)
print(y1)
# add
x1 + y1
torch.add(x1, y1)

result = torch.empty(5,3)
torch.add(x1, y1, out=result)
print(result)  # 把输出作为一个变量

# In-place operation
y1.add_(x1)  # 把操作保存在y1里面
print(y1)
# 任何in-place运算都会以_结尾。  x.copy_(y)   x.t_()会改变x

# 各种Numpy的indexing都可以在Pytorch tensor上使用
print(y1[:, 1:])  # 把所有行留下，把第一列之后的留下，相当于第零列舍去
print(y1[1:, 1:])  # 舍弃第零行，第零列

# 如果希望resize一个tensor，可以使用torch.view
x7 = torch.randn(4,4)
y2 = x7.view(16)  # 变成16维
y3 = x7.view(2,8)  # 2x8 matrix
y3 = x7.view(2,-1)  # 会自动算出对应的为数，16/2 = 8， 但不能写两个-1
# 要能被16整除，因此出现（-1， 5）会报错

# 若只有一个元素的tensor，使用.item()可以把里面的value变成python数值
x8 = torch.randn(1)
print(x8.data)  # 仍返回一个tensor
print(x8.grad)  # 返回一个grad
print(x8.item())  # 返回一个数字
print(y3.transpose(1, 0))  # 将y3进行转置

# 在Numpy和Tensor之间转换
# Torch Tensor 和 Numpy Array 共享内存，改变其中一项另一项也改变
a = torch.ones(5)
b = a.numpy()
b[1] = 2
print(a)

# 把Numpy ndarry转成Torch Tensor
c = np.ones(5)
d = torch.from_numpy(c)
np.add(c, 1, out = c)
print(c)
print(d)

# CUDA Tensors
if torch.cuda.is_available():
    device = torch.device("cuda")          # a CUDA device object
    y = torch.ones_like(x7, device=device)  # directly create a tensor on GPU
    x7 = x7.to(device)                       # or just use strings ``.to("cuda")``
    z = x7 + y
    print(z)
    print(z.to("cpu", torch.double))       # ``.to`` can also change dtype together!

# numpy是在CPU上操作的
# y.to("cpu").data.numpy()
# y.cpu().data.numpy()

使用Numpy实现两层模型

'''
用numpy实现两层神经网络，一个隐藏层，没有bias，用来从x预测y，使用L2 loss
h = W_1X + b_1
a = max(0,h)
y_hat = w_2a + b_2

numpy ndarray 是一个普通的n维array
'''
import numpy as np

N, D_in, H, D_out = 64, 1000, 100, 10  # 输入64个变量，输入是1000维，输出10维，中间层H为100维

# 随机创建一些训练数据
x = np.random.randn(N, D_in)
y = np.random.randn(N, D_out)

w1 = np.random.randn(D_in, H)
w2 = np.random.randn(H, D_out)

learning_rate = 1e-6
for t in range(500):  # forward pass
    h = x.dot(w1)    # N*H  点积
    h_relu = np.maximum(h, 0)  # N*H
    y_pred = h_relu.dot(w2)  # N*D_out

    # compute loss
    loss = np.square(y_pred - y).sum()
    print(t, loss)

    # backward pass, compute the gradient
    grad_y_pred = 2.0*(y_pred - y)
    grad_w2 = h_relu.T.dot(grad_y_pred)
    grad_h_relu = grad_y_pred.dot(w2.T)
    grad_h = grad_h_relu.copy()
    grad_h[h<0] = 0
    grad_w1 = x.T.dot(grad_h)

    # update weights of w1 and w2
    w1 -= learning_rate*grad_w1
    w2 -= learning_rate*grad_w2

使用pytorch实现两层模型

手动实现反向传播及更新

import torch

N, D_in, H, D_out = 64, 1000, 100, 10  # 输入64个变量，输入是1000维，输出10维，中间层H为100维

# 随机创建一些训练数据
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)

w1 = torch.randn(D_in, H)
w2 = torch.randn(H, D_out)

learning_rate = 1e-6
for t in range(500):  # forward pass
    h = x.mm(w1)    # N*H  matrix multipulication点积
    h_relu = h.clamp(min=0)  # N*H  类似于夹子，把值夹在min和max之间
    y_pred = h_relu.mm(w2)  # N*D_out

    # compute loss
    loss = (y_pred - y).pow(2).sum().item()  # 要转成数字
    print(t, loss)

    # backward pass, compute the gradient
    grad_y_pred = 2.0*(y_pred - y)
    grad_w2 = h_relu.t().mm(grad_y_pred)
    grad_h_relu = grad_y_pred.mm(w2.T)
    grad_h = grad_h_relu.clone()
    grad_h[h<0] = 0
    grad_w1 = x.t().mm(grad_h)

    # update weights of w1 and w2
    w1 -= learning_rate*grad_w1
    w2 -= learning_rate*grad_w2

自动实现反向传播

import torch

N, D_in, H, D_out = 64, 1000, 100, 10  # 输入64个变量，输入是1000维，输出10维，中间层H为100维

# 随机创建一些训练数据
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)

w1 = torch.randn(D_in, H, requires_grad=True)
w2 = torch.randn(H, D_out, requires_grad=True)

learning_rate = 1e-6
for t in range(500):  # forward pass
    # h = x.mm(w1)    # N*H  matrix multipulication点积
    # h_relu = h.clamp(min=0)  # N*H  类似于夹子，把值夹在min和max之间
    y_pred = x.mm(w1).clamp(min=0).mm(w2)  # N*D_out

    # compute loss
    loss = (y_pred - y).pow(2).sum()  #computation graph
    print(t, loss.item())

    # backward pass, compute the gradient
    loss.backward()

    # update weights of w1 and w2
    # 为了不让计算图占内存，不会记住w1和w2的值
    with torch.no_grad():
        w1 -= learning_rate*w1.grad
        w2 -= learning_rate*w2.grad
        w1.grad.zero_()  # 避免多次计算累加导致错误
        w2.grad.zero_()

使用pytorch的nn库实现两层网络

'''
用nn库来构建网络 neural network
用autograd来构建计算图和计算gradients
'''
import torch
import torch.nn as nn

N, D_in, H, D_out = 64, 1000, 100, 10  # 输入64个变量，输入是1000维，输出10维，中间层H为100维

# 随机创建一些训练数据
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)

model = torch.nn.Sequential(
    torch.nn.Linear(D_in, H),  # w_1*x + b_1
    torch.nn.ReLU(),
    torch.nn.Linear(H, D_out)
)

# 把初始化变成normal distribution会让模型效果好很多
torch.nn.init.normal_(model[0].weight)
torch.nn.init.normal_(model[2].weight)

# model = model.cuda()

loss_fn = nn.MSELoss(reduction='sum')

learning_rate = 1e-6
for t in range(500):  # forward pass
    y_pred = model(x)  # model.forward()

    # compute loss
    loss = loss_fn(y_pred, y)  # computation graph
    print(t, loss.item())

    model.zero_grad()  # 将梯度清零避免叠加

    # backward pass, compute the gradient
    loss.backward()

    # update weights of w1 and w2
    with torch.no_grad():
        for param in model.parameters():
            param -= learning_rate*param.grad

使用optim进行自动优化

'''
用nn库来构建网络 neural network
用autograd来构建计算图和计算gradients
'''
import torch
import torch.nn as nn

N, D_in, H, D_out = 64, 1000, 100, 10  # 输入64个变量，输入是1000维，输出10维，中间层H为100维

# 随机创建一些训练数据
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)

model = torch.nn.Sequential(
    torch.nn.Linear(D_in, H),  # w_1*x + b_1
    torch.nn.ReLU(),
    torch.nn.Linear(H, D_out)
)

# model = model.cuda()

loss_fn = nn.MSELoss(reduction='sum')
learning_rate = 1e-4
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# Adam 的学习率一般在1e-3到1e-4
# 若用SGD，则需要把初始值做一下nomalization，不知道为什么，但是loss会变得很小，玄学

for t in range(500):  # forward pass
    y_pred = model(x)  # model.forward()

    # compute loss
    loss = loss_fn(y_pred, y)  # computation graph
    print(t, loss.item())

    optimizer.zero_grad()  # 将梯度清零避免叠加

    # backward pass, compute the gradient
    loss.backward()

    # update model parameters
    optimizer.step()  # optimizer会更新

使用自定义神经网络

'''
用nn库来构建网络 neural network
用autograd来构建计算图和计算gradients
'''
import torch
import torch.nn as nn

N, D_in, H, D_out = 64, 1000, 100, 10  # 输入64个变量，输入是1000维，输出10维，中间层H为100维

# 随机创建一些训练数据
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)


# 把所有的module写在__init__里面，把每一个有导数的层放在init里面，在init里面定义模型的框架
class TwoLayerNet(torch.nn.Module):
    def __init__(self, D_in, H, D_out):
        super(TwoLayerNet, self).__init__()
        self.linear1 = torch.nn.Linear(D_in, H, bias=False)
        self.linear2 = torch.nn.Linear(H, D_out, bias=False)

    def forward(self, x):  # 前向传播的过程
        y_pred = self.linear2(self.linear1(x).clamp(min=0))
        return y_pred


model = TwoLayerNet(D_in, H, D_out)

loss_fn = nn.MSELoss(reduction='sum')
learning_rate = 1e-4
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# Adam 的学习率一般在1e-3到1e-4
# 若用SGD，则需要把初始值做一下nomalization，不知道为什么，但是loss会变得很小，玄学

for t in range(500):  # forward pass
    y_pred = model(x)  # model.forward()

    # compute loss
    loss = loss_fn(y_pred, y)  # computation graph
    print(t, loss.item())

    optimizer.zero_grad()  # 将梯度清零避免叠加

    # backward pass, compute the gradient
    loss.backward()

    # update model parameters
    optimizer.step()  # optimizer会更新