README.md

Classifying custom image datasets by creating Convolutional Neural Networks and Residual Networks from scratch with PyTorch

• Dataset

EuroSAT: Land Use and Land Cover Classification with Sentinel-2

• Create data and dataloaders

train_dataset = datasets.ImageFolder(os.path.join(DATASET_PATH, 'train'), data_transforms)
test_dataset = datasets.ImageFolder(os.path.join(DATASET_PATH, 'test'), data_transforms)

train_loader = DataLoader(train_dataset, BATCH_SIZE, shuffle=True, num_workers=2, pin_memory=True)
test_loader = DataLoader(test_dataset, BATCH_SIZE, shuffle=False, num_workers=2, pin_memory=True)

• Create transformations

from torchvision import datasets, models, transforms
data_transforms = transforms.Compose([
        transforms.Resize(64),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])

• Create models

# Create Convolutional Neural Network
model = CNN()

# Create Residual Network
model = ResNet(ResidualBlock, [2, 2, 2])

• Open and run train_cnn.ipynb and train_resnet.ipynb jupyter notebooks for train and evaluate models for Euro SAT dataset.

# Main loop
train_loss = []
train_accuracy = []
test_loss = []
test_accuracy = []
epochs = []

for epoch in range(1, NUM_EPOCHS+1):
    print(f'\n\nRunning epoch {epoch} of {NUM_EPOCHS}...\n')
    epochs.append(epoch)

    #-------------------------Train-------------------------
    
    #Reset these below variables to 0 at the begining of every epoch
    correct = 0
    iterations = 0
    iter_loss = 0.0
    
    model.train()  # Put the network into training mode
    
    for i, (inputs, labels) in enumerate(train_loader):
       
        if USE_CUDA:
            inputs = inputs.cuda()
            labels = labels.cuda()        
            
        outputs = model(inputs)
        loss = criterion(outputs, labels)
        iter_loss += loss.item()  # Accumulate the loss
        optimizer.zero_grad() # Clear off the gradient in (w = w - gradient)
        loss.backward()   # Backpropagation 
        optimizer.step()  # Update the weights
        
        # Record the correct predictions for training data 
        _, predicted = torch.max(outputs, 1)
        correct += (predicted == labels).sum()
        iterations += 1
        
    scheduler.step()
        
    # Record the training loss
    train_loss.append(iter_loss/iterations)
    # Record the training accuracy
    train_accuracy.append((100 * correct / len(train_dataset)))   
     
    #-------------------------Test--------------------------
    
    correct = 0
    iterations = 0
    testing_loss = 0.0
    
    model.eval()  # Put the network into evaluation mode
    
    for i, (inputs, labels) in enumerate(test_loader):

        if USE_CUDA:
            inputs = inputs.cuda()
            labels = labels.cuda()
        
        outputs = model(inputs)     
        loss = criterion(outputs, labels) # Calculate the loss
        testing_loss += loss.item()
        # Record the correct predictions for training data
        _, predicted = torch.max(outputs, 1)
        correct += (predicted == labels).sum()
        
        iterations += 1

    # Record the Testing loss
    test_loss.append(testing_loss/iterations)
    # Record the Testing accuracy
    test_accuracy.append((100 * correct / len(test_dataset)))
   
    print(f'\nEpoch {epoch} validation results: Loss={test_loss[-1]} | Accuracy={test_accuracy[-1]}\n')