Pytorch神經網絡擬合sin(x)函數

Pytorch簡介

PyTorch是一個基於Torch的Python開源機器學習庫，用於自然語言處理等應用程序。它主要由Facebookd的人工智能小組開發，不僅能夠 實現強大的GPU加速，同時還支持動態神經網絡，這一點是現在很多主流框架如TensorFlow都不支持的。 PyTorch提供了兩個高級功能：

• 具有強大的GPU加速的張量計算（如Numpy）
• 包含自動求導系統的深度神經網絡

用到的包

import torch
import matplotlib.pyplot as plt
import torch.nn as nn

生成訓練用的數據

x = torch.linspace(-torch.pi,torch.pi,10000) #(1000, )
x = torch.unsqueeze(input=x, dim=1)  # (1000, 1)
y = torch.sin(x)  # (1000, 1)

如果可以的話，放到GPU里訓練（快很多）

device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using {device} device to train")

x = x.to(device)
y = y.to(device)

定義模型

兩層，x --> (1,70) --> sigmoid --> (70,1) --> y_pred

class NeuralNetwork(nn.Module):
    def __init__(self):
        # 調用
        super(NeuralNetwork, self).__init__()
        #
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(1, 70),
            nn.Sigmoid(),
            nn.Linear(70,1)
        )
        
    def forward(self, x):
        y_pred = self.linear_relu_stack(x)
        return y_pred

定義損失函數和優化器，開始訓練

# 把模型放到GPU上訓練  
model = NeuralNetwork().to(device) 
#均方差做損失函數
loss_fn = nn.MSELoss()  
#optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
#用下面這個Adam優化器會收斂的快很多
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

# 迭代3000次
batches = 3000

plt.figure("regression")                     #新建一張畫布，打印數據點和預測值
plt.ion() #開啟交互模式
plt.show()
for i in range(batches):
    y_pred = model(x)
    loss = loss_fn(y_pred, y)
    
    # 
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    
    if i % 100 == 0:
        loss, batch = loss.item(), i
        print(f'loss: {loss} {batch}')
        plt.cla()
        plt.plot(x.cpu().numpy(), y.cpu().numpy())
        plt.plot(x.cpu().numpy(), y_pred.detach().cpu().numpy())
        plt.pause(0.001)
        

保存模型和讀取模型參數

# 保存
torch.save(model.state_dict(), "model.pth")
print("Saved PyTorch Model State to model.pth")
# 讀取
model = NeuralNetwork()
model.load_state_dict(torch.load("model.pth"))

結果

全部代碼

import torch
import matplotlib.pyplot as plt
import torch.nn as nn


# 生成訓練數據用的數據
x = torch.linspace(-torch.pi,torch.pi,10000) #(1000, )
x = torch.unsqueeze(input=x, dim=1)  # (1000, 1)
y = torch.sin(x)  # (1000, 1)

plt.plot(x.numpy(),y.numpy())


# 如果可以用cuda就在cuda上運行，這樣會快很多
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using {device} device to train")

x = x.to(device)
y = y.to(device)

# 定義NN模型，繼承自nn.Module
class NeuralNetwork(nn.Module):
    def __init__(self):
        # 調用
        super(NeuralNetwork, self).__init__()
        #
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(1, 70),
            nn.Sigmoid(),
            nn.Linear(70,1)
        )
        
    def forward(self, x):
        y_pred = self.linear_relu_stack(x)
        return y_pred
        

# 把模型放到GPU上訓練  
model = NeuralNetwork().to(device) 
#均方差做損失函數
loss_fn = nn.MSELoss()  
#optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
#用下面這個Adam優化器會收斂的快很多
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

# 迭代3000次
batches = 3000

plt.figure("regression")                     #新建一張畫布，打印數據點和預測值
plt.ion() #開啟交互模式
plt.show()
for i in range(batches):
    y_pred = model(x)
    loss = loss_fn(y_pred, y)
    
    # 
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    
    if i % 100 == 0:
        loss, batch = loss.item(), i
        print(f'loss: {loss} {batch}')
        plt.cla()
        plt.plot(x.cpu().numpy(), y.cpu().numpy())
        plt.plot(x.cpu().numpy(), y_pred.detach().cpu().numpy())
        plt.pause(0.001)
# 保存
torch.save(model.state_dict(), "model.pth")
print("Saved PyTorch Model State to model.pth")

import torch
import matplotlib.pyplot as plt
import torch.nn as nn


class NeuralNetwork(nn.Module):
    def __init__(self):
        # 調用
        super(NeuralNetwork, self).__init__()
        #
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(1, 70),
            nn.Sigmoid(),
            nn.Linear(70,1)
        )
        
    def forward(self, x):
        y_pred = self.linear_relu_stack(x)
        return y_pred

x = torch.linspace(-torch.pi,torch.pi,10000) #(1000, )
x = torch.unsqueeze(input=x, dim=1)  # (1000, 1)
        
model = NeuralNetwork()
model.load_state_dict(torch.load("model.pth"))

y = model(x)

plt.plot(x.numpy(), y.detach().numpy())

參考資料

1. https://pytorch.org/tutorials/beginner/basics/quickstart_tutorial.html
2. https://blog.csdn.net/weixin_41582053/article/details/109412050