Training image classifiers

Deep Learning intermedio con PyTorch

Michal Oleszak

Machine Learning Engineer

Data augmentation revisited

Image of a cat.

Data augmentation revisited

Image of a cat flipped horizontally and rotated.

What should not be augmented

Yellow lemon and green lime look the same, only differ in color.

What should not be augmented

Yellow lemon and green lime look the same, only differ in color.

What should not be augmented

The letter "W" after vertical flip looks like the letter "M".

Augmentations can impact the label
Whether this is confusing depends on the task
Always choose augmentations with the data and task in mind!

Augmentations for cloud classification

A sample of cloud pictures.

Random rotation: expose model to different angles of cloud formations
Horizontal flip: simulate different viewpoints of the sky
Auto contrast adjustment: simulate different lighting conditions

train_transforms = transforms.Compose([
    transforms.RandomHorizontalFlip(),
    transforms.RandomRotation(45),
    transforms.RandomAutocontrast(),
    transforms.ToTensor(),
    transforms.Resize((128, 128))
])

Cross-Entropy loss

Binary classification: binary cross-entropy (BCE) loss
Multi-class classification: cross-entropy loss
criterion = nn.CrossEntropyLoss()

Image classifier training loop

net = Net(num_classes=7)

criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=0.001)


for epoch in range(10):
    for images, labels in dataloader_train:
        optimizer.zero_grad()
        outputs = net(images)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

Let's practice!

Deep Learning intermedio con PyTorch