Image Classification

Module: Computer Vision | Difficulty: Beginner

ResNet Skip Connections

Enables training of very deep networks (100+ layers).

EfficientNet Scaling

Compound scaling: depth , width , resolution

subject to

Data Augmentation

• Random horizontal flip
• Random crop with padding
• Color jittering
• Cutout: zero out random patches
• Mixup: blend two images:

Transfer Learning

Fine-tune last layers with small learning rate.

import torch
import torch.nn as nn
import torchvision.models as models

def create_classifier(num_classes, pretrained=True):
    model = models.efficientnet_b0(pretrained=pretrained)
    model.classifier[1] = nn.Linear(model.classifier[1].in_features, num_classes)
    return model

class Cutout:
    def __init__(self, length):
        self.length = length
    def __call__(self, img):
        h, w = img.shape[1:], img.shape[2:]
        mask = torch.ones_like(img)
        cy = torch.randint(h, (1,))
        cx = torch.randint(w, (1,))
        y1 = (cy - self.length // 2).clamp(0, h)
        y2 = (cy + self.length // 2).clamp(0, h)
        x1 = (cx - self.length // 2).clamp(0, w)
        x2 = (cx + self.length // 2).clamp(0, w)
        mask[:, y1:y2, x1:x2] = 0
        return img * mask

Key Takeaways

• Residual connections enable very deep network training
• EfficientNet provides optimal accuracy-efficiency trade-off
• Transfer learning dramatically reduces data requirements