পাঠ ১২ · ৩৫-এর মধ্যে · মডিউল ২

Home / AI Courses / Computer Vision / VGG

VGG — গভীর কিন্তু সরল

VGG — deep but simple, all 3×3 convolutions

৭ মিনিট পড়া মাঝারি · Intermediate PyTorch কোডসহ

এই পাঠে যা শিখবেন

VGG-এর uniform architecture philosophy
3×3 conv stacking-এর গাণিতিক যুক্তি
VGG-16 ও VGG-19 layer-by-layer
VGG-এর modern uses

১ · VGG — Oxford-এর contribution

Karen Simonyan ও Andrew Zisserman, Oxford University Visual Geometry Group (VGG). ICLR 2015 paper. ILSVRC 2014-এ classification 2nd place (1st GoogLeNet).

VGG-এর philosophy: "Just go deeper, with the simplest possible building block."

কেন্দ্রীয় ধারণা

Architecture diversity-র বদলে uniformity. প্রতিটি conv 3×3, প্রতিটি pool 2×2. Depth-ই key. সরলতাই beauty.

২ · 3×3 conv stacking-এর math

AlexNet 11×11 ও 5×5 use করেছিল। VGG বলল — না, পরপর 3×3 conv ব্যবহার, একই RF কিন্তু ভাল।

RF calculation:

একটি 7×7 conv: RF = 7.
তিনটি 3×3 conv: RF = $3 + 2 + 2 = 7$।
RF সমান।

Parameter comparison (single channel):

7×7 conv: 49 params.
তিনটি 3×3: $3 \times 9 = 27$ params.
45% কম।

Non-linearity:

7×7 + ReLU: 1 ReLU.
তিনটি 3×3 + ReLU each: 3 ReLU.
Decision boundary বেশি expressive.

৩ · VGG-16 architecture

Input: 224×224×3.

Block 1: Conv64 (3×3) × 2 → MaxPool (2×2) → 112×112×64.
Block 2: Conv128 × 2 → Pool → 56×56×128.
Block 3: Conv256 × 3 → Pool → 28×28×256.
Block 4: Conv512 × 3 → Pool → 14×14×512.
Block 5: Conv512 × 3 → Pool → 7×7×512.
FC layers: 4096 → 4096 → 1000.

Pattern: spatial halve every block, channel double (64 → 128 → 256 → 512 → 512).

VGG-19: Block 3, 4, 5-এ চারটি conv (extra one). ~144M parameters.

৪ · কেন এই uniformity?

Reasoning simplicity: একটা rule — সব ক্ষেত্রে apply.
Search space: hyperparameter কমানো — kernel size fix.
Hardware efficiency: 3×3 conv well-optimized in cuDNN.
Reproducibility: অন্য researcher easy implement.

VGG = "Lego block" approach — একই block বারবার stack করে tower বানানো। AlexNet ছিল mixed (11×11, 5×5, 3×3). VGG-র uniform approach পরে ResNet, EfficientNet-এ inherit.

৫ · VGG-এর সমস্যা

Parameter heavy: 138M — bulk মূলত FC layer-এ।
Memory: training-এ activation 12-14 GB.
Slow: deep + dense FC.
Vanishing gradient: 19+ layer-এ train difficult — ResNet-এর জন্ম এই সমস্যা থেকে।
Single-stream sequential: কোনো branching বা skip — limited representational power.

VGG-16 — ৫ Conv block + ৩ FC. Uniform 3×3 conv ও 2×2 pool — সরলতার মধ্যে গভীরতা।

৬ · PyTorch-এ VGG

Python · PyTorch

import torch
from torchvision.models import vgg16, VGG16_Weights

# Pretrained VGG-16
model = vgg16(weights=VGG16_Weights.IMAGENET1K_V1)
model.eval()

x = torch.randn(1, 3, 224, 224)
with torch.no_grad():
    out = model(x)
print("Output:", out.shape)            # (1, 1000)

# Parameter
n = sum(p.numel() for p in model.parameters())
print(f"Parameters: {n:,}")            # ~138M

# Feature extractor (without classifier)
features = model.features              # Sequential conv blocks
with torch.no_grad():
    feat = features(x)
print("Feature map:", feat.shape)      # (1, 512, 7, 7)

VGG-এর features অংশ — ImageNet-এ pretrained backbone. Style transfer, perceptual loss, ও ছোট dataset transfer learning-এ এই part-ই use হয়।

৭ · VGG-এর modern legacy

Perceptual loss: super-resolution, image generation-এ — VGG feature distance loss.
Style transfer: Gatys et al. (২০১৫) — VGG-19-এর Gram matrix.
SSIM/LPIPS metric: VGG-feature-based perceptual similarity.
Pre-trained backbone: ছোট dataset detection-এ Faster R-CNN VGG-16.
Pedagogical: "first deep CNN" tutorial-এ।

৮ · VGG vs পরবর্তী

VGG vs ResNet-50: RN-50 25M params (5x কম), 76% top-1 (3% better), 4x faster.
VGG vs EfficientNet-B0: EN-B0 5M params, 77% top-1, mobile-runnable.
VGG vs ConvNeXt: CN-T 28M, 82%, modern design.

৯ · VGG-এর শিক্ষা

Simplicity scales: uniform pattern + depth = good performance.
Small kernel preferred: 3×3 modern CNN-এর default.
Double channels at downsample: ResNet, EfficientNet-এ continue.
Pre-trained backbone: transfer learning-এর philosophy.

VGG আজ "old" — কিন্তু its philosophy ResNet, ConvNeXt, MobileNet-এ embedded. 3×3 conv-এর dominance VGG থেকে।

ভাবনার প্রশ্ন

প্র ০১ VGG-19 (১৯ layer) VGG-16-এর তুলনায় সামান্য better, কিন্তু VGG-22, 26 — সরাসরি deeper করলে accuracy কমে। কেন? এই observation-ই কি ResNet-এর জন্ম?

এটি DL-এর সবচেয়ে impactful empirical observation.

Degradation problem:

VGG-16: 71.5% top-1.
VGG-19: 71.8%.
Hypothetical VGG-30: 70% (worse!)।
Counter-intuitive — deeper should be better (more capacity).

Cause analysis:

Vanishing gradient: backprop chain rule — deep layer-এ tiny gradient.
Optimization difficulty: deep network-এর loss landscape rugged.
Identity mapping hard: network যদি last layer "pass-through" শেখাতে চায়, weight tune করতে কঠিন।

Critical insight (He et al., 2015):

"যদি deep network-এর extra layer identity হত — তাহলে accuracy অন্তত equal থাকতো।"
"যেহেতু worse, optimization fail করছে identity শেখাতে।"
"সমাধান: identity-কে easy default বানানো।"

ResNet-এর solution:

$$y = F(x) + x$$

Skip connection — input directly add.
$F(x) = 0$ → identity (default).
Network only learns "residual".
Gradient skip-এ flow — vanishing solve.

Result:

ResNet-152 (১৫২ layer) — train possible.
VGG-style 152 layer — diverge.
ImageNet 2015 winner: ResNet 3.57% error (vs VGG 7.3%).

Subsequent insights:

BatchNorm — independent fix to vanishing gradient.
Highway Networks (gated skip) — slightly earlier.
DenseNet — every layer connected to all previous.

Modern best practice:

Deep network-এ skip connection essential.
Transformer-এ residual connection — same idea.
EfficientNet, ConvNeXt — সব use করে।

মূল উপলব্ধি: VGG-র degradation-ই DL-এর সবচেয়ে impactful empirical paradox. সমাধান (skip connection) — architecture design-এর central principle.

প্র ০২ VGG-এর FC layer-এ 102M parameters. FC remove করে GAP দিয়ে replace করলে accuracy কতটা impacted? Modern ResNet তো এটাই করে।

এটি network-in-network paper (Lin et al., 2013) থেকে শুরু architecture trend.

VGG FC vs GAP:

VGG FC1: 7×7×512=25088 → 4096 = 102M params.
GAP-equivalent: 7×7 spatial average → 512 vector → 1000 = 0.5M params.
200x reduction.

Accuracy impact:

"VGG-GAP" — 71.5% → 70% (1.5% drop).
Negligible loss compared to enormous parameter saving.
Modern architectures — GAP standard.

কেন GAP কাজ করে:

Final feature map already abstract — spatial average meaningful.
Implicit regularization — each class probability spatial average.
"Class activation map" naturally — interpretable.

FC-এর role যা GAP-এ লস:

Spatial position-aware classification — যদি object position matter.
Complex feature combination.
আজ — এই extra capacity রক্ষায় conv depth বাড়ানো better.

Variants:

GAP + linear: ResNet, MobileNet.
GAP + small FC (256-512): EfficientNet.
GMP (max): some networks.
GAP + GMP concat: detail + summary.
Adaptive pool: variable input size.

Input size flexibility:

VGG: fixed 224×224 — FC requires fixed dim.
GAP: any size — adaptive avg pool spatial dim collapse.
Detection/segmentation-এ huge benefit.

Modern hybrid:

ConvNeXt: GAP + LayerNorm + linear.
Swin Transformer: similar.
ViT: CLS token (different paradigm).

মূল উপলব্ধি: "Big FC = bad" আজকের consensus. VGG-এর historical importance — এটি দেখাল FC-এর problem. Modern minimalist heads industry standard.

প্র ০৩ VGG-19-এর Gram matrix style transfer-এর foundation. Gram matrix ঠিক কী, কেন style capture করে?

Gatys, Ecker & Bethge (২০১৫) — neural style transfer. Internet-এ viral. VGG-এর role essential.

Gram matrix definition:

Feature map $F \in \mathbb{R}^{C \times HW}$ (channel × spatial flatten).
Gram matrix $G = F F^T \in \mathbb{R}^{C \times C}$।
Element $G_{ij} = \sum_k F_{ik} F_{jk}$ — channel $i$ ও $j$-এর correlation.

Spatial information removal:

Gram = inner product across spatial location.
Feature কোথায় (where) discard, কী (what) preserve.
Texture-এ "where" matter করে না — "what pattern" matters.

Style = correlation pattern:

Brushstroke — local texture pattern.
VGG layer-এ — channel-i "edge" detect, channel-j "color blob" detect.
Their co-occurrence (correlation) = style signature.
Van Gogh-এর painting-এ specific pattern co-activation.

Style transfer algorithm:

Content image $C$, style image $S$, output $O$।
Content loss: $\| F^L(O) - F^L(C) \|^2$ — high-level VGG feature match.
Style loss: $\sum_l \| G^l(O) - G^l(S) \|^2$ — Gram matrix match across layer.
$O$ initialize random, optimize via gradient descent.

VGG-এর role:

Pretrained on ImageNet — diverse natural pattern.
Hierarchical features — low-level texture থেকে high-level object.
Layer-wise style match → multi-scale texture transfer.

Modern alternative:

Fast NST (Johnson et al.): feedforward network train — realtime.
AdaIN (Huang & Belongie): adaptive instance normalization.
Stable Diffusion: text-guided style transfer.
VGG-এর role-ও সরে যাচ্ছে — কিন্তু perceptual loss-এ এখনো standard.

Bangladesh-এ application:

Local artist (S.M. Sultan, Zainul Abedin) style transfer app.
Cultural heritage digitization.
Educational tool — art history visualize.

মূল উপলব্ধি: Mathematical elegance — Gram matrix-এর simplicity-এ texture-এর semantic meaning. CV-র creative application-এর gateway.

প্র ০৪ VGG-16 ২০১৪-এর। ১২ বছর পর — কোনো production system এ এখনো VGG ব্যবহার করার valid reason আছে?

Pragmatically — কিছু niche-এ VGG এখনো relevant.

Cases যেখানে VGG পছন্দ:

Perceptual loss: super-resolution, image-to-image translation. VGG feature matching standard.
Style transfer: Gatys-style — VGG-19 still gold.
SSIM/LPIPS: perceptual quality metric — VGG-based.
Legacy systems: pre-2018 production — migration cost.
Educational: simplicity teaches CNN concepts.

Cases যেখানে VGG NOT okay:

Inference latency critical: mobile, edge — MobileNet, EfficientNet better.
Memory limited: 138M parameters massive.
Modern accuracy bar: VGG 71% vs ConvNeXt 84% — 13% gap.
Detection/segmentation: ResNet/EfficientNet backbone better.

Specific applications still using VGG:

SRGAN, ESRGAN — super-resolution generator+discriminator perceptual loss.
Pix2Pix, CycleGAN — feature matching loss.
GauGAN, StyleGAN — analysis network.
Few medical research codebase — historical.

Modern alternative for similar role:

Perceptual loss: EfficientNet, CLIP feature also work.
LPIPS update: AlexNet/VGG/SqueezeNet variants.
DINO feature: self-supervised, often better.

Bangladesh context:

Limited GPU access — VGG inference slow.
EfficientNet-B0 better starting point.
Research tutorial-এ VGG OK for clarity.

Migration recommendation:

Start: ResNet-50 (general).
Mobile: MobileNetV3, EfficientNet-Lite.
Accuracy: ConvNeXt, EfficientNetV2.
Foundation: CLIP, DINOv2 features.

মূল উপলব্ধি: Architecture aging — like programming language. VGG-এর design principle alive (3×3, deep, simple). Specific weights-এর use case narrow. Engineer-এর judgment crucial.

অনুশীলন

Param count: VGG-16 Block 1-এর প্রথম conv (3 → 64, 3×3, padding=1). Param কত?

$3 \times 64 \times 3 \times 3 + 64 = 1{,}792$।
RF check: VGG Block 5-এর শেষ conv-এর approximate receptive field?

Each pool RF doubles. 5 pool 32x downsample. Conv-এর accumulated effect — RF প্রায় ২১২×২১২ — ছবির বেশিরভাগ। তাই global context captured.
ভাবুন: VGG-16 vs MobileNet-V3 — accuracy কাছাকাছি কিন্তু MobileNet 25x ছোট। Mobile inference-এ কোনটা?

অবশ্যই MobileNet-V3. 5M params, ARM-friendly depthwise separable conv, optimized squeeze-and-excitation. Real device-এ 10-20x faster.