<aside> ðŸ›
Deep Residual Networks (ResNet) ImageNet Classification
Classifies images into 1,000 ImageNet categories using residual learning with skip connections. Three variants available: ResNet-18 (69.57% top-1), ResNet-50 (75.99% top-1), ResNet-101 (77.56% top-1). Trade-off accuracy for speed/parameters.
</aside>
ResNet classifies images into 1,000 ImageNet categories using residual learning (skip connections). Architecture enables training of very deep networks (18–101 layers) without vanishing gradient problems. Three model sizes available for different accuracy/efficiency requirements.
Input: RGB image
Output: Logits for 1,000 ImageNet classes
| Model | Top-1 Accuracy | Top-5 Accuracy | Parameters |
|---|---|---|---|
| ResNet-18 | 69.57% | 89.24% | 11.7M |
| ResNet-50 | 75.99% | 92.98% | 25.6M |
| ResNet-101 | 77.56% | 93.79% | 44.5M |
Best For:
| Architecture | Residual Learning with Skip Connections |
|---|---|
| Core Innovation: y = F(x) + x | • F(x): Residual function learned by stacked layers |
| • x: Skip connection (identity bypass) | |
| • Benefits: Enables gradient flow through deep networks, mitigates vanishing gradient | |
| Building block | • ResNet-18: Basic blocks (2 conv layers per block) |
| • ResNet-50/101: Bottleneck blocks (1×1 reduce → 3×3 → 1×1 restore) | |
| Training | • Dataset: ImageNet-1k (1.28M training images, 1,000 classes) |
| • Optimizer: SGD (momentum 0.9) | |
| • Batch size: 256 | |
| • Learning rate: 0.1 (÷10 every 30 epochs) | |
| • Epochs: 100 | |
| • Weight decay: 1e-4 | |
| • Data augmentation: |
◦ Random crop 224×224
â—¦ Random horizontal flip
◦ Multi-scale jitter (256–480)
â—¦ ImageNet normalization (per-channel mean/std) |
| Depth Per Stage | • ResNet-18: [2, 2, 2, 2] • ResNet-50: [3, 4, 6, 3] • ResNet-101: [3, 4, 23, 3] |