2nd/10_Wiki/Topics/AI_and_ML/ResNet-Architectures.md

---
id: wiki-2026-0508-resnet-architectures
title: ResNet Architectures
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [ResNet, Residual Networks, Skip Connection]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [resnet, cnn, skip-connection, deep-learning, vision]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pytorch
---

# ResNet Architectures

## 매 한 줄
> **"매 identity shortcut 의 deep network 의 가능"**. ResNet (He et al. 2015) 의 skip connection 으로 152-layer training 의 enable, ImageNet 우승. 2026 의 ConvNeXt-V2/Hiera 의 ResNet idea 의 Vision Transformer 의 hybrid.

## 매 핵심

### 매 핵심 idea
- **Residual block**: `y = F(x) + x` — 매 identity 의 default.
- **Vanishing gradient 의 mitigate**: gradient 의 skip 의 통해 직접 flow.
- **Deeper = better** (until 1000+ where diminishing).
- **Bottleneck (1x1 → 3x3 → 1x1)**: ResNet-50+ 의 efficiency.

### 매 Variants
- **ResNet-18 / 34**: basic block (no bottleneck).
- **ResNet-50 / 101 / 152**: bottleneck — 매 default backbone.
- **Wide ResNet**: wider, shallower.
- **ResNeXt**: grouped conv (cardinality).
- **DenseNet**: 매 모든 prev layer 의 concat (vs sum).
- **ConvNeXt (2022) / V2 (2023)**: ResNet 의 ViT-style modernize — depthwise conv, LayerNorm, GELU, 매 ImageNet SOTA 의 ViT 의 match.
- **Hiera (Meta, 2023)**: hierarchical ViT, ResNet 의 spirit.

### 매 응용
1. Image classification backbone (ImageNet, medical imaging).
2. Object detection (Faster R-CNN, RetinaNet 의 ResNet backbone).
3. Segmentation (U-Net + ResNet encoder).
4. Feature extraction for downstream (CLIP image encoder origin).
5. Diffusion model U-Net (residual 의 everywhere).

## 💻 패턴

### Residual block (PyTorch)
```python
import torch.nn as nn

class BasicBlock(nn.Module):
    expansion = 1
    def __init__(self, in_c, out_c, stride=1):
        super().__init__()
        self.conv1 = nn.Conv2d(in_c, out_c, 3, stride, 1, bias=False)
        self.bn1   = nn.BatchNorm2d(out_c)
        self.conv2 = nn.Conv2d(out_c, out_c, 3, 1, 1, bias=False)
        self.bn2   = nn.BatchNorm2d(out_c)
        self.shortcut = nn.Identity()
        if stride != 1 or in_c != out_c:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_c, out_c, 1, stride, bias=False),
                nn.BatchNorm2d(out_c),
            )
    def forward(self, x):
        out = torch.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        out = out + self.shortcut(x)        # 매 핵심
        return torch.relu(out)
```

### Bottleneck block (ResNet-50+)
```python
class Bottleneck(nn.Module):
    expansion = 4
    def __init__(self, in_c, mid_c, stride=1):
        super().__init__()
        out_c = mid_c * self.expansion
        self.conv1 = nn.Conv2d(in_c, mid_c, 1, bias=False)
        self.bn1   = nn.BatchNorm2d(mid_c)
        self.conv2 = nn.Conv2d(mid_c, mid_c, 3, stride, 1, bias=False)
        self.bn2   = nn.BatchNorm2d(mid_c)
        self.conv3 = nn.Conv2d(mid_c, out_c, 1, bias=False)
        self.bn3   = nn.BatchNorm2d(out_c)
        self.shortcut = nn.Identity()
        if stride != 1 or in_c != out_c:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_c, out_c, 1, stride, bias=False),
                nn.BatchNorm2d(out_c))
    def forward(self, x):
        out = torch.relu(self.bn1(self.conv1(x)))
        out = torch.relu(self.bn2(self.conv2(out)))
        out = self.bn3(self.conv3(out))
        return torch.relu(out + self.shortcut(x))
```

### Pretrained ResNet-50 (torchvision)
```python
import torch
from torchvision.models import resnet50, ResNet50_Weights

weights = ResNet50_Weights.IMAGENET1K_V2
model = resnet50(weights=weights).eval().cuda()
preprocess = weights.transforms()

img = preprocess(load_image("cat.jpg")).unsqueeze(0).cuda()
with torch.no_grad():
    logits = model(img)
print(weights.meta["categories"][logits.argmax(1).item()])
```

### ConvNeXt (2026 modern alt)
```python
from torchvision.models import convnext_base, ConvNeXt_Base_Weights
model = convnext_base(weights=ConvNeXt_Base_Weights.IMAGENET1K_V1)
# 매 ResNet 의 spirit, ViT-grade accuracy
```

### Fine-tune for custom task
```python
model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
for p in model.parameters(): p.requires_grad = False
model.fc = nn.Linear(model.fc.in_features, num_classes)  # train only head
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Quick CV baseline | ResNet-50 pretrained |
| SOTA accuracy | ConvNeXt-V2 / ViT-L / Hiera |
| Edge / mobile | MobileNetV3 / EfficientNet-Lite |
| Detection backbone | ResNet-50 + FPN, ConvNeXt for SOTA |
| Diffusion U-Net | Residual blocks (ResNet-style) |

**기본값**: ResNet-50 의 baseline, ConvNeXt-Base 의 production target.

## 🔗 Graph
- 부모: [[CNN]] · [[Deep-Learning]]
- 응용: [[Image-Classification-Mastery]] · [[Object-Detection]] · [[Diffusion-Models]]
- Adjacent: [[Skip-Connection]]

## 🤖 LLM 활용
**언제**: ResNet implementation 의 explain, paper summarization (He 2015), debugging gradient flow.
**언제 X**: actual training (use PyTorch + GPU), benchmark numbers (verify on Papers with Code).

## ❌ 안티패턴
- **No skip connection 매 deep**: 매 50+ layers 의 vanishing gradient.
- **BatchNorm 의 small batch**: <16 의 broken — GroupNorm/LayerNorm 의 use.
- **Train from scratch 매 small data**: 매 pretrain 의 always.
- **Skip connection 의 add 의 다른 shape**: 매 1x1 conv projection 의 needed.
- **ResNet-152 매 mobile**: 60M params — MobileNet/EfficientNet 의 use.

## 🧪 검증 / 중복
- Verified (He et al. 2015 ResNet, Liu et al. 2022 ConvNeXt, torchvision docs).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — ResNet 매 ConvNeXt revival 의 connect |