GAN — Generative Adversarial Network

Mode collapse — GAN research-এর central pain point। Causes multifaceted।

Mathematical perspective:

• $G$ optimal — point distribution-এ collapse চাইতে পারে।
• $D$ fixed হলে $G$ argmax $\log D(G(z))$ — এক mode যথেষ্ট।
• Diversity-এর কোনো explicit incentive নেই objective-এ।

Game dynamics:

• $G$ একটি mode পায় যা D fool করে।
• D সেই mode reject শিখে।
• $G$ আরেকটি mode-এ jump।
• Cyclical — never coverage।

Loss landscape:

• JS-divergence — disjoint distribution-এ saturate।
• Gradient signal poor।
• $G$ gradient escape difficult।

Fixes — by category:

(১) Loss redesign:

• WGAN — Wasserstein metric, smooth gradient সর্বত্র।
• Hinge loss — robust।
• Least squares GAN।

(২) Architecture:

• Spectral normalization — Lipschitz $D$।
• Self-attention — global pattern।
• Progressive growing।

(৩) Training tricks:

• Minibatch discrimination — D batch-এ diversity check।
• Feature matching — generated features match real distribution।
• Unrolled GAN — D-এর future steps predict।
• Two time-scale update rule (TTUR)।

(৪) Optimization:

• EMA — running average of G weights।
• R1, R2 regularization।
• Lower learning rate।

Empirical observation:

• StyleGAN2 + mode-coverage techniques — significant improvement।
• Diversity metric (LPIPS) তবু fragile।
• Class-conditional GAN — partial fix।

Current standing:

• Diffusion — mode coverage natural।
• GAN-এ ongoing technique stack।
• StyleGAN3 — texture-stable, mode-stable।
• Diffusion taking over generation।

মূল উপলব্ধি: Mode collapse — GAN objective-এ implicit। Diversity না engineered, automatic নয়। Wasserstein + tricks-এর stack lessen। Diffusion fundamentally different — mode coverage natural। GAN art + science — production-এ caution।

WGAN — GAN history-এর landmark। Mathematical foundation শক্ত।

Why Lipschitz:

• Wasserstein distance dual form — $\sup_{f \in \text{Lip-1}} \mathbb{E}[f(x)] - \mathbb{E}[f(y)]$।
• $D$ ($f$) — Lipschitz-1 constraint।
• Gradient bounded by ১ everywhere।
• Smooth loss landscape।

Original WGAN — weight clipping:

• $D$-এর প্রতিটি weight $[-c, c]$-এ clip।
• Crude approximation Lipschitz-এর।
• $c = 0.01$ typical।

Problems:

• Weight distribution bimodal — most at $\pm c$।
• Capacity loss — D underutilized।
• Gradient explode/vanish — depending $c$।
• Hyperparameter sensitive।

WGAN-GP — gradient penalty (২০১৭):

• Real-fake interpolated point $\hat{x}$।
• $D$-এর gradient $\hat{x}$-এ — penalty $(\|\nabla\| - 1)^2$।
• Lipschitz-1 push exactly।

Implementation:

def gradient_penalty(D, real, fake):
    eps = torch.rand(real.size(0), 1, 1, 1)
    interp = eps * real + (1-eps) * fake
    interp.requires_grad_(True)
    d_interp = D(interp)
    grad = torch.autograd.grad(
        d_interp.sum(), interp, create_graph=True)[0]
    grad = grad.view(grad.size(0), -1)
    return ((grad.norm(2, dim=1) - 1) ** 2).mean()

Comparison:

Weight clipping:

• Simple, fast।
• Capacity loss।
• Hyperparameter brittle।

Gradient penalty:

• More principled।
• Computationally costly (extra grad)।
• BatchNorm-এর সাথে interaction issue (use LayerNorm/InstanceNorm in D)।

Spectral normalization (২০১৮):

• Each layer-এর spectral norm clip।
• Lipschitz layer-wise।
• Computationally efficient।
• SAGAN, BigGAN-এ ব্যবহৃত।

Practical recommendation:

• SN-GAN — modern default।
• WGAN-GP — research standard।
• Original WGAN — historical।

Production:

• StyleGAN — R1 penalty (gradient on real only)।
• Lower compute, similar effect।
• Spectral norm + R1 — strong combination।

মূল উপলব্ধি: Lipschitz constraint Wasserstein-এর mathematical requirement। Weight clipping ad-hoc, gradient penalty principled, spectral norm efficient। R1 modern production। GAN training Lipschitz enforcement — central engineering challenge।

StyleGAN-এর architecture insight — image generation-এর landmark contribution।

Resolution hierarchy:

• Coarse layers (4×4 → 8×8) — pose, face shape, hair style।
• Middle layers (16×16 → 32×32) — facial features, eyes, nose।
• Fine layers (64×64+) — color, texture, lighting।
• Spatial scale → semantic scale natural correspondence।

Mapping network:

• $z \to w$ — 8-layer MLP।
• $z$ Gaussian, $w$ disentangled।
• $\mathcal{W}$-space "linear separable"।

AdaIN injection:

• $w$ — affine projection per layer।
• Layer activation normalize, scale-shift।
• Per-layer style control।

Style mixing trick:

• Two latent $w_1, w_2$।
• Layer 1-3 use $w_1$, layer 4-end use $w_2$।
• $w_1$ identity, $w_2$ texture — mixed image।

Why it works:

• Mapping network — disentanglement encourage।
• Style mixing regularization — independent layer।
• Mixing during training — robustness।
• $w$-space-এ semantic direction discoverable।

Disentanglement test:

• "Smile" direction — $w$-এ identify।
• Add direction → smile add।
• Other attribute unchanged।
• Linear interpretability।

Editing application:

• GAN inversion — image → $w$।
• $w$-এ edit।
• Decode → edited image।
• FaceShop, deepfake editing।

StyleGAN2 improvements:

• Modulated conv — AdaIN replacement।
• Path length regularization।
• Lazy regularization।
• Photorealistic ১০২৪×১০২৪।

StyleGAN3:

• Texture sticking fix।
• Equivariance to translation/rotation।
• Smooth interpolation।

Compared to alternatives:

• VAE — less disentangled।
• BigGAN — class-conditioned, less interpretable।
• Diffusion — different mechanism, prompt-based control।

Bangla face editing:

• StyleGAN fine-tune — Bangla face dataset।
• Cultural attribute (age, attire) editing।
• Limited dataset — challenge।
• Transfer learning critical।

মূল উপলব্ধি: StyleGAN — architectural innovation enable disentanglement। Spatial-semantic correspondence emergent। $w$-space linear control। Modern face generation gold standard। Editing applications powerful।

Practical Bangladesh use case — handwritten Bangla character generation। Data augmentation-এর crucial application।

Use case:

• OCR system trained on synthetic + real।
• Bangla handwritten data scarce।
• Style variation needed।
• Production inference fast।

Option 1 — VAE:

Pros:

• Stable training।
• Latent interpolation smooth।
• Probabilistic foundation।

Cons:

• Blurry — handwriting sharpness lost।
• OCR training-এ poor signal।
• Unsuitable।

Option 2 — GAN (DCGAN/StyleGAN):

Pros:

• Sharp output — handwriting essential।
• Style mixing — diverse strokes।
• Fast inference।
• Established for character generation (HWGAN)।

Cons:

• Training instability।
• Mode collapse — character variety risk।
• Hyperparameter tuning art।

Option 3 — Diffusion:

Pros:

• Best quality।
• Mode coverage natural।
• Prompt-based control (class-conditional)।

Cons:

• Slow inference (diffusion step)।
• Heavier compute training।
• Recent — fewer Bangla-specific resource।

Recommendation — phased approach:

Phase 1 — Conditional GAN:

• Class label condition — 'ক', 'খ', etc।
• BanglaLekha-Isolated dataset (Hossain et al.)।
• ~৮০ class, ~১৫০K image।
• DCGAN baseline।

Phase 2 — StyleGAN-style:

• Style mixing — different writer style।
• Coarse style → letter shape, fine → stroke।
• Diverse handwriting।

Phase 3 — Diffusion (if budget allows):

• Best quality production।
• Latent diffusion in compressed space।
• OCR training boost।

Implementation:

# Conditional DCGAN for Bangla characters
class CondGenerator(nn.Module):
    def __init__(self, z_dim=100, n_class=80, img_dim=784):
        super().__init__()
        self.label_emb = nn.Embedding(n_class, n_class)
        self.net = nn.Sequential(
            nn.Linear(z_dim + n_class, 256),
            nn.LeakyReLU(0.2),
            nn.Linear(256, 512), nn.LeakyReLU(0.2),
            nn.Linear(512, img_dim), nn.Tanh())
    def forward(self, z, y):
        c = self.label_emb(y)
        return self.net(torch.cat([z, c], dim=1))

Evaluation:

• FID score (vs real handwriting)।
• OCR accuracy boost — primary metric।
• Human evaluation — readability।
• Diversity metric — style variation।

Production deployment:

• One-time bulk generation।
• Variation per session।
• Cloud GPU — minimal cost।
• Augmentation pipeline integration।

Bangla handwriting challenges:

• Conjunct character ("যুক্তাক্ষর") complexity।
• Vowel mark positioning ("কার")।
• Cursive variation।
• Modifier consistency।

Hybrid approach:

• Real data + synthetic 50-50 mix।
• OCR train на combined।
• Test on real held-out।
• Improvement validate।

Bangladesh practical:

• OCR critical — government, banking, education।
• Bangla data scarcity bottleneck।
• Synthetic augmentation high value।
• GAN-based feasible practical solution।

মূল উপলব্ধি: Synthetic Bangla handwriting — GAN best practical। StyleGAN style variation। Diffusion quality but expensive। Phased approach: GAN → StyleGAN → Diffusion। OCR augmentation impact significant। Bangladesh data-scarcity solution generation-based।