AlexNet.py

import numpy as np
from datetime import datetime 

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader

from torchvision import datasets, transforms

import matplotlib.pyplot as plt

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

# parameters
RANDOM_SEED = 42
LEARNING_RATE = 0.005
BATCH_SIZE = 64
N_EPOCHS = 15

IMG_SIZE = 32
N_CLASSES = 10

def train(train_loader, model, criterion, optimizer, device):
    '''
    Function for the training step of the training loop
    '''

    model.train()
    running_loss = 0
    for X, y_true in train_loader:
        optimizer.zero_grad()
        
        X = X.to(device)
        y_true = y_true.to(device)
    
        # Forward pass
        y_hat, _ = model(X) 
        loss = criterion(y_hat, y_true) 
        running_loss += loss.item() * X.size(0)

        # Backward pass
        loss.backward()
        optimizer.step()
        
    epoch_loss = running_loss / len(train_loader.dataset)
    return model, optimizer, epoch_loss

def validate(valid_loader, model, criterion, device):
    '''
    Function for the validation step of the training loop
    '''
   
    model.eval()
    running_loss = 0
    
    for X, y_true in valid_loader:
    
        X = X.to(device)
        y_true = y_true.to(device)

        # Forward pass and record loss
        y_hat, _ = model(X) 
        loss = criterion(y_hat, y_true) 
        running_loss += loss.item() * X.size(0)

    epoch_loss = running_loss / len(valid_loader.dataset)
        
    return model, epoch_loss
    
def get_accuracy(model, data_loader, device):
    '''
    Function for computing the accuracy of the predictions over the entire data_loader
    '''
    
    correct_pred = 0 
    n = 0
    
    with torch.no_grad():
        model.eval()
        for X, y_true in data_loader:

            X = X.to(device)
            y_true = y_true.to(device)

            _, y_prob = model(X)
            _, predicted_labels = torch.max(y_prob, 1)

            n += y_true.size(0)
            correct_pred += (predicted_labels == y_true).sum()

    return correct_pred.float() / n

def plot_losses(train_losses, valid_losses):
    '''
    Function for plotting training and validation losses
    '''
    
    # temporarily change the style of the plots to seaborn 
    plt.style.use('seaborn')

    train_losses = np.array(train_losses) 
    valid_losses = np.array(valid_losses)

    fig, ax = plt.subplots(figsize = (8, 4.5))

    ax.plot(train_losses, color='blue', label='Training loss') 
    ax.plot(valid_losses, color='red', label='Validation loss')
    ax.set(title="Loss over epochs", 
            xlabel='Epoch',
            ylabel='Loss') 
    ax.legend()
    fig.show()
    
    # change the plot style to default
    plt.style.use('default')

def training_loop(model, criterion, optimizer, train_loader, valid_loader, epochs, device, print_every=1):
    '''
    Function defining the entire training loop
    '''
    
    # set objects for storing metrics
    best_loss = 1e10
    train_losses = []
    valid_losses = []
 
    # Train model
    for epoch in range(0, epochs):

        # training
        model, optimizer, train_loss = train(train_loader, model, criterion, optimizer, device)
        train_losses.append(train_loss)

        # validation
        with torch.no_grad():
            model, valid_loss = validate(valid_loader, model, criterion, device)
            valid_losses.append(valid_loss)

        if epoch % print_every == (print_every - 1):
            
            train_acc = get_accuracy(model, train_loader, device=device)
            valid_acc = get_accuracy(model, valid_loader, device=device)
                
            print(f'{datetime.now().time().replace(microsecond=0)} --- '
                  f'Epoch: {epoch}\t'
                  f'Train loss: {train_loss:.4f}\t'
                  f'Valid loss: {valid_loss:.4f}\t'
                  f'Train accuracy: {100 * train_acc:.2f}\t'
                  f'Valid accuracy: {100 * valid_acc:.2f}')

    plot_losses(train_losses, valid_losses)
    
    return model, optimizer, (train_losses, valid_losses)

# define transforms
transforms = transforms.Compose([transforms.Resize((224, 224)),
                                 transforms.ToTensor()])

# download and create datasets
train_dataset = datasets.MNIST(root='mnist_data', 
                               train=True, 
                               transform=transforms,
                               download=True)

valid_dataset = datasets.MNIST(root='mnist_data', 
                               train=False, 
                               transform=transforms)

# define the data loaders
train_loader = DataLoader(dataset=train_dataset, 
                          batch_size=BATCH_SIZE, 
                          shuffle=True)

valid_loader = DataLoader(dataset=valid_dataset, 
                          batch_size=BATCH_SIZE, 
                          shuffle=False)

class AlexNet(nn.Module):
  def __init__(self):
    super(AlexNet,self).__init__()
    
    self.feature_extractor = nn.Sequential(
        #First Layer
        nn.Conv2d(in_channels=1, out_channels=96, kernel_size=11, stride=4),
        nn.ReLU(),
        nn.MaxPool2d(kernel_size=3, stride=2),

        #Second Layer
        nn.Conv2d(in_channels=96, out_channels=256, kernel_size=5, stride=1, padding=2),
        nn.ReLU(),
        nn.MaxPool2d(kernel_size=3, stride=2),

        #Third Layer
        nn.Conv2d(in_channels=256, out_channels=384, kernel_size=3, stride=1, padding=1),
        nn.ReLU(),
        nn.Conv2d(in_channels=384, out_channels=384, kernel_size=3, stride=1, padding=1),
        nn.ReLU(),
        nn.Conv2d(in_channels=384, out_channels=256, kernel_size=3, stride=1, padding=1),
        nn.ReLU(),
        nn.MaxPool2d(kernel_size=3, stride=2),
    )

    self.cl1 = nn.Linear(5*5*256, 4096)
    self.cl2 = nn.Linear(4096, 4096)
    self.cl3 = nn.Linear(4096, 10)
  
  def forward(self, x):
    x = self.feature_extractor(x)
    x = torch.flatten(x,1)

    x = F.relu(self.cl1(x))
    x = F.dropout(x,0.5)
    x = F.relu(self.cl2(x))
    x = F.dropout(x,0.5)
    logits = self.cl3(x)
    prob = F.log_softmax(logits, dim=1)

    return logits, prob

torch.manual_seed(RANDOM_SEED)

model = AlexNet().to(DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)
criterion = nn.CrossEntropyLoss()

model, optimizer, _ = training_loop(model, criterion, optimizer, train_loader, valid_loader, N_EPOCHS, DEVICE)

torch.save(model, 'dat.pth')