পাঠ ১৬ · ৩৫-এর মধ্যে · মডিউল ২
Home / AI Courses / Computer Vision / Transfer learning

Transfer learning ব্যবহারিক

Transfer learning — practical fine-tuning of pretrained CNNs
৭ মিনিট পড়া মাঝারি · Intermediate PyTorch কোডসহ

এই পাঠে যা শিখবেন

  • Transfer learning কেন কাজ করে
  • Feature extraction vs fine-tuning
  • Layer freezing strategy
  • Learning rate scheduling — discriminative LR

১ · কেন transfer learning?

Practical CV-তে — ImageNet-এর মতো ১.৪M labeled image কেউ collect করে না। আপনার Bangla street sign dataset হয়তো ১০০০। Scratch থেকে train করলে — overfit.

Transfer learningTransfer Learningএকটি task-এ trained model-এর knowledge অন্য task-এ ব্যবহার। CV-তে ImageNet-এ pretrained CNN আপনার specific task-এ adapt — gold standard. idea: ImageNet-এ trained model edge, texture, shape — universal feature শিখেছে। আপনার task-এ এই foundation reuse.

কেন্দ্রীয় ধারণা

Early layer features (edge, color) universal — যেকোনো image task-এ relevant. Late layer features task-specific. Transfer = early layer reuse, late layer retrain.

২ · তিনটি strategy

(ক) Feature extraction

  • সব layer freeze (no gradient).
  • শুধু final classifier (FC) train.
  • Use case: very small dataset (<1K), domain ImageNet-এর কাছাকাছি।

(খ) Fine-tuning (full)

  • সব layer trainable, কিন্তু low learning rate.
  • Use case: medium dataset (5K-50K), domain similar.
  • Risk: pretrained weight destroy যদি LR বেশি।

(গ) Gradual unfreeze (discriminative LR)

  • প্রথমে head only train (epoch 1-3).
  • তারপর — last block unfreeze, lower LR.
  • ক্রমশ deeper unfreeze.
  • Use case: large dataset, domain ImageNet থেকে far.

৩ · কখন কোন strategy

Dataset sizeDomain similarityStrategy
<1KSimilarFeature extraction
<1KDifferentFeature extract + augment + simpler model
1K-10KSimilarHead + fine-tune last block
10K+SimilarFull fine-tune, low LR
10K+DifferentGradual unfreeze, discriminative LR
100K+Very differentPretrain initialization, full retrain

৪ · ImageNet pretrained model পাওয়া

  • torchvision.models — ResNet, EfficientNet, ViT, ConvNeXt.
  • timm (Python lib): 1000+ models — Ross Wightman maintained.
  • Hugging Face Hub: বহু custom-trained model.
  • OpenCLIP, DINOv2: self-supervised foundation models.
Transfer learning = একজন experienced rangoli artist-কে Bengali alpana শেখানো। তিনি color, symmetry, brush technique জানেন — শুধু specific Bengali pattern শেখাতে হবে। Scratch — child থেকে শুরু, অসম্ভব।
Transfer learning — তিন strategy Pretrained ResNet-50 Conv1+Pool Block1 (3 res) Block2 (4 res) Block3 (6 res) Block4 (3 res) FC 1000 A) Feature extract 🔒 Frozen 🔒 Frozen 🔒 Frozen 🔒 Frozen 🔒 Frozen ✏️ New head B) Gradual unfreeze 🔒 LR 0 🔒 LR 0 LR 1e-5 LR 1e-4 LR 1e-3 ✏️ LR 1e-2 C) Full FT All trainable low LR 1e-4 unified across ✏️ Train
Transfer learning — frozen feature থেকে gradual unfreeze থেকে full fine-tune. Dataset size + domain similarity-এর উপর choice.

৫ · PyTorch implementation

Python · PyTorch 
import torch
import torch.nn as nn
from torchvision.models import resnet50, ResNet50_Weights

# Pretrained ResNet-50
model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)

# Strategy A — feature extraction
def setup_feature_extract(model, num_classes=2):
    for p in model.parameters():
        p.requires_grad = False
    # Replace classifier head
    model.fc = nn.Linear(model.fc.in_features, num_classes)
    return model

# Strategy B — gradual unfreeze
def setup_gradual(model, num_classes=2):
    # Freeze all
    for p in model.parameters():
        p.requires_grad = False
    # Unfreeze last block + head
    for p in model.layer4.parameters():
        p.requires_grad = True
    model.fc = nn.Linear(model.fc.in_features, num_classes)
    return model

# Discriminative LR optimizer
def get_optimizer(model, head_lr=1e-3, body_lr=1e-4):
    return torch.optim.Adam([
        {'params': model.layer1.parameters(), 'lr': body_lr * 0.1},
        {'params': model.layer2.parameters(), 'lr': body_lr * 0.3},
        {'params': model.layer3.parameters(), 'lr': body_lr},
        {'params': model.layer4.parameters(), 'lr': body_lr * 3},
        {'params': model.fc.parameters(),     'lr': head_lr},
    ])

m = setup_gradual(model, num_classes=10)
opt = get_optimizer(m)
print("Trainable params:", sum(p.numel() for p in m.parameters() if p.requires_grad))

    


    

      Discriminative LR — early layer-এ low LR, head-এ high. ImageNet feature subtle preserve, new head fast adapt.
    


    
৬ · Practical tips

    
    
      
  • Same preprocessing: pretraining-এর mean/std normalization same use.
    • 
      
  • Input size: 224×224 ImageNet standard. ভিন্ন size — adaptive pool.
    • 
      
  • Batch size: 32-128 typical. BatchNorm sensitive — খুব ছোট না।
    • 
      
  • BN-এ eval mode (inference): running stats use, not batch.
    • 
      
  • Freeze BN-ও: small dataset-এ — BN running stat freeze (otherwise drift).
    • 
      
  • Augmentation always: small dataset-এ heavy augmentation.
    • 
    


    
৭ · Domain shift

    
    
      
  • Natural image: ImageNet pretrain ideal.
    • 
      
  • Medical imaging: partial transfer ভালো — early layer. Domain-specific pretrain (RadImageNet) better.
    • 
      
  • Satellite/aerial: rotation augmentation ভিন্ন domain — extra fine-tune.
    • 
      
  • Document/OCR: ImageNet limited transfer — text image very different.
    • 
      
  • Microscopy: domain pretrain (BiomedCLIP) better.
    • 
    


    
৮ · কী model choose?

    
    
      
  • Quick baseline: ResNet-18.
    • 
      
  • Strong baseline: ResNet-50, EfficientNet-B0.
    • 
      
  • Mobile target: MobileNetV3, EfficientNet-B0.
    • 
      
  • Maximum accuracy: ConvNeXt, EfficientNet-B5.
    • 
      
  • Dense prediction (segmentation): ResNet-50 backbone + U-Net decoder.
    • 
    


    
৯ · Beyond ImageNet

    
    
      
  • CLIP: 400M image-text pair. Zero-shot classification possible.
    • 
      
  • DINOv2: 1.4B images, self-supervised. SOTA features.
    • 
      
  • SAM: Segment anything — segmentation foundation.
    • 
      
  • RadImageNet: medical imaging-specific.
    • 
      
  • iNaturalist: nature/wildlife images.
    • 
    


    

      Transfer learning-এর danger — domain mismatch. ImageNet (natural photo) থেকে satellite/medical-এ — early layer transferable, late layer mismatch. Always validate on held-out test.
    


  


  

    
    
  


  

    
 ভাবনার প্রশ্ন


    

      
        প্র ০১
        Transfer learning ImageNet pretrain থেকে — আজকের CLIP, DINOv2 foundation models কীভাবে এই paradigm বদলে দিচ্ছে?
      
      

        

          ২০২২-পরবর্তী era-এ "transfer learning"-এর অর্থ ক্রমশ বদলাচ্ছে।
        

        

          ImageNet pretrain era:
        

        
    
          
  • Supervised, 1.4M label.
    • 
          
  • 1000-class bias.
    • 
          
  • Natural photo domain.
    • 
          
  • End task: fine-tune head.
    • 
        

        

          Foundation model era:
        

        
    
          
  • CLIP (2021, OpenAI): 400M image-text pair. Zero-shot classification.
    • 
          
  • DINOv2 (2023, Meta): 1.4B images, self-supervised. Feature linear-probe SOTA.
    • 
          
  • SAM (2023, Meta): 11M images, 1B masks. Segmentation foundation.
    • 
          
  • EVA-02 (2023): SOTA on 28+ task.
    • 
        

        

          Paradigm shift:
        

        
    
          
  • Model size: 60M → 1B+.
    • 
          
  • Data: 1M → 1B+.
    • 
          
  • Pretrain method: supervised → self-supervised.
    • 
          
  • Use mode: fine-tune → linear probe / zero-shot.
    • 
        

        

          Zero-shot CLIP example:
        

        
import clip
model, preprocess = clip.load("ViT-L/14")
image = preprocess(img).unsqueeze(0)
texts = ["a photo of a cat", "a photo of a dog"]
text_tokens = clip.tokenize(texts)
# Zero-shot — no fine-tune needed
similarity = model(image, text_tokens)

        

          Implications:
        

        
    
          
  • Less data needed: few-shot or zero-shot work.
    • 
          
  • Better transfer: diverse pretrain → robust.
    • 
          
  • Multimodal: vision + language together.
    • 
          
  • API-mediated: finetune complex, prompt-engineer instead.
    • 
        

        

          Linear probe vs fine-tune:
        

        
    
          
  • Linear probe — frozen features, train linear classifier.
    • 
          
  • DINOv2 linear probe ImageNet 86%.
    • 
          
  • Often competitive with full fine-tune.
    • 
          
  • 1000x faster, no catastrophic forgetting.
    • 
        

        

          Bangladesh deployment:
        

        
    
          
  • CLIP-based — Bangla text not always supported (multilingual variant দরকার)।
    • 
          
  • DINOv2 features — universal, language-independent.
    • 
          
  • Bangla-specific finetune cost lower.
    • 
        

        

          Practical recommendation 2026:
        

        
    
          
  • Start: linear probe DINOv2.
    • 
          
  • If linear probe insufficient — full fine-tune.
    • 
          
  • Mobile — distill DINOv2 to smaller.
    • 
          
  • Multi-modal — CLIP/SigLIP.
    • 
        

        

          মূল উপলব্ধি: "Transfer learning" → "foundation model adaptation". Same idea, different scale. ImageNet pretrain — first generation. Foundation model — second. Multimodal generalist — third.
        

      

    


    

      
        প্র ০২
        "Catastrophic forgetting" — fine-tuning-এর সময় pretrain knowledge হারিয়ে যাওয়া। কীভাবে rotate করবেন?
      
      

        

          এটি transfer learning-এর সবচেয়ে underestimated risk. Aggressive fine-tune-এ pretrained features destroy.
        

        

          Catastrophic forgetting symptom:
        

        
    
          
  • Train accuracy high, but ImageNet-style images-এ model fail.
    • 
          
  • Domain-specific overfitting — generalize ভাঙে।
    • 
          
  • Linear probe accuracy drop after fine-tune.
    • 
        

        

          Causes:
        

        
    
          
  • High learning rate — pretrained weight wipe.
    • 
          
  • Long training — overfit small dataset.
    • 
          
  • Aggressive augmentation — distribution shift.
    • 
          
  • BatchNorm running stats drift.
    • 
        

        

          Mitigations:
        

        
    
          
  • Low LR: 1e-4 to 1e-5 (vs 1e-3 scratch).
    • 
          
  • Discriminative LR: early layer slower.
    • 
          
  • Early stopping: validation accuracy plateau-এ stop.
    • 
          
  • Limited unfreezing: last 2-3 block only.
    • 
          
  • Weight decay: regularize toward zero.
    • 
          
  • Frozen BN: running stats keep.
    • 
        

        

          Advanced techniques:
        

        
    
          
  • Elastic Weight Consolidation (EWC): important weight regularize.
    • 
          
  • Knowledge distillation: student match teacher (pretrained) on auxiliary data.
    • 
          
  • LoRA (Low-Rank Adaptation): small additional weights, original frozen.
    • 
          
  • Adapter layers: small task-specific layer insert.
    • 
          
  • Prompt tuning: input prompt learn, model frozen.
    • 
        

        

          LoRA detailed:
        

        
    
          
  • $W' = W + \Delta W$, $\Delta W = AB$ (rank $r$)।
    • 
          
  • $r \ll d$ — much fewer parameter.
    • 
          
  • Original $W$ frozen.
    • 
          
  • Originally for LLM, now CV-তেও popular.
    • 
          
  • Easy to swap multiple task-specific LoRAs.
    • 
        

        

          Multi-task fine-tune:
        

        
    
          
  • Original ImageNet objective + new task — joint loss.
    • 
          
  • Forgetting prevent.
    • 
          
  • Computational cost বেশি।
    • 
        

        

          Validation strategy:
        

        
    
          
  • Two test set:
            
      
              
    • (a) Target domain (your task).
      • 
              
    • (b) Source domain (ImageNet sample).
      • 
            
    
          
    • 
          
  • (b) accuracy যদি drop — forgetting evidence.
    • 
        

        

          মূল উপলব্ধি: Fine-tune ≠ free. Pretrained weight valuable knowledge. Preserve consciously — discriminative LR, regularization, gentle adaptation.
        

      

    


    

      
        প্র ০৩
        Bangla street sign detect করতে চান — 200 labeled image. কী strategy? Domain difference কতটুকু matter?
      
      

        

          এটি real Bangladesh CV problem. Hidayat-er ola, Pathao-র route assist — এই ধরনের use case.
        

        

          Dataset profile:
        

        
    
          
  • 200 labeled image — খুব ছোট।
    • 
          
  • Bangla text + symbol — ImageNet-এ rare.
    • 
          
  • Outdoor lighting variability.
    • 
          
  • Mobile camera quality.
    • 
        

        

          Step-by-step approach:
        

        
    
          
  1. Augmentation aggressive: rotation ±10°, brightness ±25%, perspective, weather (rain/fog) augment. 200 → effective 5000+.
    2. 
          
  2. Pretrained model: EfficientNet-B0 ImageNet OR DINOv2-base.
    3. 
          
  3. Strategy: feature extract + small head:
            
      
              
    • Freeze backbone.
      • 
              
    • Add small MLP (256 → num_classes).
      • 
              
    • Train with heavy regularization.
      • 
            
    
          
    4. 
          
  4. If insufficient: unfreeze last block, very low LR (1e-5).
    5. 
        

        

          Domain-specific concerns:
        

        
    
          
  • Bangla character: ImageNet-এ rare.
    • 
          
  • Solution: synthetic data — rendered Bangla sign on natural background.
    • 
          
  • Lighting: outdoor + tropical — augment heavy.
    • 
          
  • Occlusion: trees, vehicles, dust — Random Erasing/CutOut.
    • 
          
  • Camera quality: low-res augment.
    • 
        

        

          Better than ImageNet-pretrain:
        

        
    
          
  • OpenStreetMap: sign images public.
    • 
          
  • Mapillary: street-view image dataset.
    • 
          
  • SVHN, GTSRB: closer domain (street sign).
    • 
          
  • Pretrain on these (intermediate) → fine-tune Bangla.
    • 
        

        

          Synthetic data approach:
        

        
    
          
  • Bangla font render on backgrounds.
    • 
          
  • 3D simulation (Blender, Unity).
    • 
          
  • 10K+ synthetic + 200 real → mix train.
    • 
          
  • Stable Diffusion image generation — stress test.
    • 
        

        

          Active learning:
        

        
    
          
  • Initial 200 → train baseline.
    • 
          
  • Predict on 10K unlabeled — identify "hard" cases (low confidence).
    • 
          
  • Manually label hard ones.
    • 
          
  • Retrain — accuracy bump.
    • 
          
  • Iterate.
    • 
        

        

          Ensemble strategy:
        

        
    
          
  • 3 models train (different seed, augment).
    • 
          
  • Average prediction.
    • 
          
  • 2-3% accuracy gain typically.
    • 
        

        

          Real-world example:
        

        
    
          
  • Bornom (Bengali OCR) — synthetic + ResNet pretrain.
    • 
          
  • BUET LICTI lab — Bangla street sign — DINOv2 finetune.
    • 
        

        

          মূল উপলব্ধি: Small data + domain-specific = creative pipeline. Pretrain + augment + synthetic + active learning — multi-pronged approach. Pure transfer learning often insufficient.
        

      

    


    

      
        প্র ০৪
        Industry-এ "ImageNet pretrain"-এর fairness ও bias problem আছে। কী bias ও কীভাবে mitigate?
      
      

        

          ImageNet famously biased — race, gender, geography. Pretrained transfer-এ bias inherited.
        

        

          Documented biases (Buolamwini & Gebru, 2018):
        

        
    
          
  • Racial: dark-skin face accuracy 35% lower.
    • 
          
  • Gender: female under-represented in many class.
    • 
          
  • Geographic: Western objects dominant. Bangladesh-এর rickshaw, vendor — minimal.
    • 
          
  • Class imbalance: 1000 dog breed, but limited "kitchen" diversity.
    • 
        

        

          Source of bias:
        

        
    
          
  • Crawled from US/Western web.
    • 
          
  • Image labeled by Mechanical Turk — mostly Western.
    • 
          
  • Search terms English-biased.
    • 
          
  • Categories WordNet — Western taxonomy.
    • 
        

        

          Bias propagation:
        

        
    
          
  • Pretrained CNN — feature space biased.
    • 
          
  • Transfer learning — bias inherited downstream.
    • 
          
  • Detection বা face analysis-এ — disparate impact.
    • 
        

        

          Mitigation strategies:
        

        
    
          
  • Diverse pretrain:
            
      
              
    • Open Images (more international).
      • 
              
    • YFCC100M (Flickr — global).
      • 
              
    • LAION-5B (web-scale, more diverse).
      • 
              
    • FAIR-PMD (Facebook — multilingual).
      • 
            
    
          
    • 
          
  • Domain-specific finetune:
            
      
              
    • Bangladesh dataset construct.
      • 
              
    • Local crowdsource labeling.
      • 
              
    • Cultural sensitivity in categories.
      • 
            
    
          
    • 
          
  • Bias auditing:
            
      
              
    • Test set stratified by demographic.
      • 
              
    • Per-group accuracy measure.
      • 
              
    • Fairness metrics (demographic parity, equal opportunity).
      • 
            
    
          
    • 
          
  • Adversarial debiasing:
            
      
              
    • Train classifier with adversary predict demographic.
      • 
              
    • Force feature non-demographic-predictive.
      • 
            
    
          
    • 
          
  • Re-weighting:
            
      
              
    • Underrepresented group weight ↑।
      • 
              
    • Loss reweight.
      • 
            
    
          
    • 
        

        

          Bangladesh-specific:
        

        
    
          
  • Diverse skin tones in South Asia — face system test.
    • 
          
  • Religious diversity (clothing, headwear).
    • 
          
  • Rural vs urban scene difference.
    • 
          
  • Bengali New Year, Eid, Durga Puja — cultural events misrecognize.
    • 
        

        

          Regulatory landscape:
        

        
    
          
  • EU AI Act — high-risk system require bias audit.
    • 
          
  • NYC Local Law 144 — automated hiring bias audit.
    • 
          
  • Bangladesh Data Protection Bill — emerging.
    • 
        

        

          Industry practice:
        

        
    
          
  • Model card — Mitchell et al. (2019).
    • 
          
  • Datasheet for dataset.
    • 
          
  • Fairness toolkit (Aequitas, AI Fairness 360).
    • 
          
  • Pre-deployment audit.
    • 
        

        

          মূল উপলব্ধি: Transfer learning convenience-এর সাথে inherited bias. Engineer-এর কাজ — bias awareness + mitigation + auditing. Bangladesh AI development-এ local context priority.
        

      

    

  


  

    
 অনুশীলন

    
    
      
  1. 
        Param freeze: ResNet-50-এর শুধু last 2 block (layer3, layer4) trainable, বাকি frozen — কোডটি লিখুন।
        
    
          
          
    
            
    for name, p in model.named_parameters():
    p.requires_grad = ('layer3' in name or 'layer4' in name or 'fc' in name)
    
          
    
        
    
      
    2. 

      
  2. 
        Strategy choose: 50K Bangla flower image, 10 species — কোন strategy?
        
    
          
          
    
            
    50K medium-large, ImageNet flower category overlap. Full fine-tune low LR (1e-4). Data augment.
    
          
    
        
    
      
    3. 

      
  3. 
        ভাবুন: Pretrained model বাছার সময় size vs accuracy trade-off — কখন EfficientNet-B7 (big) vs B0 (small)?
        
    
          
          
    
            
    B7 — server inference, accuracy critical, batch processing. B0 — mobile, edge, realtime, small dataset (B7 overfit risk).
    
          
    
        
    
      
    4. 
    

  


  

    
 আরও পড়ুন · ABCL TECH-এ আপনার পরবর্তী পদক্ষেপ

    

    

      কোড রানার কাজ না করলে? ব্রাউজারে কাজ না করলে
      Google Colab
      ব্যবহার করুন — Google-এর ফ্রি অনলাইন Python পরিবেশ, শুধু Gmail অ্যাকাউন্ট লাগে।
    

  







  
  
     এই পাঠে ভুল পেয়েছেন? জানান
  





  
    
 পূর্ববর্তী পাঠ

    
পাঠ ১৫ · DenseNet ও EfficientNet