PyTorch trains RNN, GRU, LSTM: handwritten digit recognition

pytorch neural network training demo

Dataset: MNIST

The content of this data set is handwritten digit recognition, which is divided into two parts, containing 60,000 training images and 10,000 test images respectively.

Image source: https://tensornews.cn/mnist_intro/

Neural Network: RNN, GRU, LSTM

# Imports
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import DataLoader
import torchvision.datasets as datasets
import torchvision.transforms as transforms

# Set device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Hyperparameters
input_size = 28
sequence_length = 28
num_layers = 2
hidden_size = 256
num_classes = 10
learning_rate = 0.001
batch_size = 64
num_epochs = 2

# Create a RNN
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.rnn = nn.RNN(input_size, hidden_size, num_layers, batch_first=True)
        self.fc = nn.Linear(hidden_size*sequence_length, num_classes) # fully connected
    
    def forward(self, x):
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)

        # Forward Prop
        out, _ = self.rnn(x, h0)
        out = out.reshape(out.shape[0], -1)
        out = self.fc(out)

        return out
    
# Create a GRU
class RNN_GRU(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, num_classes):
        super(RNN_GRU, self).__init__()
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.gru = nn.GRU(input_size, hidden_size, num_layers, batch_first=True)
        self.fc = nn.Linear(hidden_size*sequence_length, num_classes) # fully connected
    
    def forward(self, x):
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)

        # Forward Prop
        out, _ = self.gru(x, h0)
        out = out.reshape(out.shape[0], -1)
        out = self.fc(out)

        return out


# Create a LSTM
class RNN_LSTM(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, num_classes):
        super(RNN_LSTM, 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)
        self.fc = nn.Linear(hidden_size*sequence_length, num_classes) # fully connected
    
    def forward(self, x):
        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)

        # Forward Prop
        out, _ = self.lstm(x, (h0, c0))
        out = out.reshape(out.shape[0], -1)
        out = self.fc(out)
        return out
    

# Load data
train_dataset = datasets.MNIST(root='dataset/', 
                               train=True, 
                               transform=transforms.ToTensor(),
                               download=True)
train_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
test_dataset = datasets.MNIST(root='dataset/', 
                              train=False, 
                               transform=transforms.ToTensor(),
                               download=True)
test_loader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=True)

# Initialize network 选择一个即可
model = RNN(input_size, hidden_size, num_layers, num_classes).to(device)
# model = RNN_GRU(input_size, hidden_size, num_layers, num_classes).to(device)
# model = RNN_LSTM(input_size, hidden_size, num_layers, num_classes).to(device)

# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=learning_rate)

# Train network
for epoch in range(num_epochs):
    # data: images, targets: labels
    for batch_idx, (data, targets) in enumerate(train_loader):
        # Get data to cuda if possible
        data = data.to(device).squeeze(1) # 删除一个张量中所有维数为1的维度 (N, 1, 28, 28) -> (N, 28, 28)
        targets = targets.to(device)

        # forward
        scores = model(data) # 64*10
        loss = criterion(scores, targets)

        # backward
        optimizer.zero_grad()
        loss.backward()

        # gradient descent or adam step
        optimizer.step()


# Check accuracy on training & test to see how good our model
def check_accuracy(loader, model):
    if loader.dataset.train:
        print("Checking accuracy on training data")
    else:
        print("Checking accuracy on test data")
    num_correct = 0
    num_samples = 0
    model.eval()

    with torch.no_grad(): # 不计算梯度
        for x, y in loader:
            x = x.to(device).squeeze(1)
            y = y.to(device)
            # x = x.reshape(x.shape[0], -1) # 64*784

            scores = model(x)# 64*10
            _, predictions = scores.max(dim=1) #dim=1，表示对每行取最大值，每行代表一个样本。
            num_correct += (predictions == y).sum()
            num_samples += predictions.size(0) # 64

        print(f'Got {
      
      num_correct} / {
      
      num_samples} with accuracy {
      
      float(num_correct)/float(num_samples)*100:.2f}%')

    model.train()

check_accuracy(train_loader, model)
check_accuracy(test_loader, model)

Result

RNN Result
Checking accuracy on training data
Got 57926 / 60000 with accuracy 96.54%
Checking accuracy on test data
Got 9640 / 10000 with accuracy 96.40%


GRU Result
Checking accuracy on training data
Got 59058 / 60000 with accuracy 98.43%
Checking accuracy on test data
Got 9841 / 10000 with accuracy 98.41%

LSTM Result
Checking accuracy on training data
Got 59248 / 60000 with accuracy 98.75%
Checking accuracy on test data
Got 9849 / 10000 with accuracy 98.49%

Reference sources

【1】https://www.youtube.com/watch?v=Gl2WXLIMvKA&list=PLhhyoLH6IjfxeoooqP9rhU3HJIAVAJ3Vz&index=5