PyTorch

PyTorch 中级篇(3):循环神经网络(Recurrent Neural Network)

参考代码

yunjey的 pytorch tutorial系列

循环神经网络 学习资源

一直以为,循环神经网络使用在语音处理上的,跟我这个研究计算机视觉的没有多大关系,所以一直都回避RNN。

这里居然有RNN对MINIST数据的网络实现,那就顺带把RNN给学了。

RNN结合CNN可以用于描述照片,正好能跟计算机视觉结合起来。

介绍视频(没有原理)

什么是循环神经网络 RNN (深度学习)? What is Recurrent Neural Networks (deep learning)?

相关网页

(新手向)能否简单易懂的介绍一下RNN(循环神经网络)?

一文搞懂RNN(循环神经网络)基础篇.

【译】 理解 LSTM 网络

当然这些都是基础版本的RNN,RNN的魅力在于它的各种变化版本能用来解决各种不同形式的问题。

【图片来源】循环神经网络RNN打开手册
RNN的形式

Pytorch实现

many to one 的形式解决MINIST数据集 手写数字分类问题。

1
2
3
4
5
# 包
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
1
2
3
4
# 设备配置
# Device configuration
torch.cuda.set_device(1) # 这句用来设置pytorch在哪块GPU上运行
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
1
2
3
4
5
6
7
8
9
10
# 超参数设置
# Hyper-parameters
sequence_length = 28
input_size = 28
hidden_size = 128
num_layers = 2
num_classes = 10
batch_size = 100
num_epochs = 2
learning_rate = 0.01

MINIST 数据集

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
# 训练数据
train_dataset = torchvision.datasets.MNIST(root='../../../data/minist/',
train=True,
transform=transforms.ToTensor(),
download=True)

# 测试数据
test_dataset = torchvision.datasets.MNIST(root='../../../data/minist/',
train=False,
transform=transforms.ToTensor())

# 训练数据加载器
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)

# 测试数据加载器
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)

循环神经网络搭建

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
class RNN(nn.Module):
def __init__(self, input_size, hidden_size, num_layers, num_classes):
super(RNN, self).__init__()
self.hidden_size = hidden_size
self.num_layers = num_layers
self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True) # 选用LSTM RNN结构
self.fc = nn.Linear(hidden_size, num_classes) # 最后一层为全连接层,将隐状态转为分类

def forward(self, x):
# 初始化隐层状态和细胞状态
h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)
c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)

# 前向传播LSTM
out, _ = self.lstm(x, (h0, c0)) # 输出大小 (batch_size, seq_length, hidden_size)

# 解码最后一个时刻的隐状态
out = self.fc(out[:, -1, :])
return out
1
2
3
4
5
6
7
8
# 实例化一个模型
# 注意输入维度,虽然我不懂将一幅图28x28拆成28个大小为28的序列有啥意义
model = RNN(input_size, hidden_size, num_layers, num_classes).to(device)

# 定义损失函数和优化器
# Adam: A Method for Stochastic Optimization
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

训练模型

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
total_step = len(train_loader)
for epoch in range(num_epochs):
for i, (images, labels) in enumerate(train_loader):
images = images.reshape(-1, sequence_length, input_size).to(device) # 注意维度
labels = labels.to(device)

# 前向传播
outputs = model(images)
loss = criterion(outputs, labels)

# 反向传播和优化,注意梯度每次清零
optimizer.zero_grad()
loss.backward()
optimizer.step()

if (i+1) % 100 == 0:
print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'
.format(epoch+1, num_epochs, i+1, total_step, loss.item()))
Epoch [1/2], Step [100/600], Loss: 0.4569
Epoch [1/2], Step [200/600], Loss: 0.2823
Epoch [1/2], Step [300/600], Loss: 0.3512
Epoch [1/2], Step [400/600], Loss: 0.1702
Epoch [1/2], Step [500/600], Loss: 0.3181
Epoch [1/2], Step [600/600], Loss: 0.1821
Epoch [2/2], Step [100/600], Loss: 0.1540
Epoch [2/2], Step [200/600], Loss: 0.0848
Epoch [2/2], Step [300/600], Loss: 0.1985
Epoch [2/2], Step [400/600], Loss: 0.1537
Epoch [2/2], Step [500/600], Loss: 0.0988
Epoch [2/2], Step [600/600], Loss: 0.0315

测试模型并保存

1
2
3
4
5
6
7
8
9
10
11
12
13
# 测试集
with torch.no_grad():
correct = 0
total = 0
for images, labels in test_loader:
images = images.reshape(-1, sequence_length, input_size).to(device)
labels = labels.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()

print('Test Accuracy of the model on the 10000 test images: {} %'.format(100 * correct / total))
Test Accuracy of the model on the 10000 test images: 97.47 %
1
2
# 保存模型
torch.save(model.state_dict(), 'model.ckpt')