Oy 3 — Mashqlar to'plami

🟢 Easy

PyTorch Basics

  1. 5 ta turli shape'dagi tensor yarating va shape, dtype, device atributlarini chiqaring.
  2. requires_grad=True bilan oddiy funksiyalar uchun gradient'larni hisoblang.
  3. nn.Module subclass yarating — input → 3 ta hidden → output.

Training

  1. MNIST'da MLP 95%+ accuracy.
  2. Optimizers solishtirish: SGD, SGD+momentum, Adam, AdamW.
  3. Learning rate ni 1e-1, 1e-3, 1e-5 qilib effect ko'rish.

CNN

  1. SimpleCNN CIFAR-10'da train (5 epoch).
  2. Pretrained ResNet-18 yuklang, ImageNet rasm classifier.
  3. torchvision.transforms bilan augmentation qatorini yarating.

RNN/LSTM

  1. nn.RNN, nn.LSTM, nn.GRU ni bir xil masala uchun solishtiring.
  2. Sin function uchun next-step forecasting.
  3. Bidirectional LSTM yarating, oddiy LSTM bilan farqi.

🟡 Medium

Production-ready training

  1. Full training pipeline: Mixed precision + early stopping + checkpoint + W&B logging.
  2. Hyperparameter tuning: Optuna bilan PyTorch model uchun.
  3. Multi-GPU(Colab Pro yoki Kaggle bilan): nn.DataParallel.

Transfer learning

  1. Flower classification: 102 turdagi gullar — pretrained EfficientNet, 92%+ accuracy.
  2. Custom domain: O'zingiz tasvir to'plang (telefon kamerasi), 5 ta sinf, 50 ta rasm har sinfda — transfer learning bilan ishlatish.
  3. Few-shot learning: 5 ta rasm har sinfdan, 90%+ accuracy olishga harakat.

Time series

  1. Real stock data(yfinance): LSTM + sliding window forecasting.
  2. Multivariate: bir nechta xususiyat (price, volume, indicators) bilan LSTM.
  3. Prophet vs LSTMsolishtirish.

Text

  1. IMDB sentiment: LSTM bilan 85%+ accuracy.
  2. News classification: 4-5 ta kategoriya (AG News).
  3. Char-level language model: Shakespeare yoki o'zbek matnda.

🔴 Hard

1. Production ML API

  • Image classification (EfficientNet) FastAPI
  • Multi-stage Dockerfile (build → runtime)
  • Async batching (vakt va GPU optimization)
  • Healthcheck, metrics endpoint
  • Load test (Locust bilan): 100 req/s ga chiday oladigan optimization

2. Distributed training

  • Kaggle Notebooks Pro yoki Colab Pro
  • DistributedDataParallel bilan 2 GPU
  • Mixed precision + gradient accumulation
  • Trening vaqtini single GPU bilan solishtirish

3. Model interpretation service

  • ResNet bilan rasm classification
  • Grad-CAM ham qaytaradigan endpoint
  • Streamlit yoki React UI

4. End-to-end CV pipeline

  • Data: web'dan rasmlar to'plash (Selenium yoki API)
  • Labelling (Label Studio yoki manual)
  • Training (PyTorch + W&B)
  • Deploying (FastAPI + Docker + Nginx)
  • Monitoring (Prometheus + Grafana)

Mini-loyihalar

Mini-loyiha 1: Plant Disease Detector

  • Dataset: PlantVillage (Kaggle)
  • Transfer learning bilan 95%+ accuracy
  • Mobile-friendly (TFLite yoki PyTorch Mobile)
  • Streamlit demo

Mini-loyiha 2: Real-time Pose Estimation

  • MediaPipe yoki MMPose
  • Webcam streaming
  • WebSocket + FastAPI

Mini-loyiha 3: Music Genre Classifier

  • GTZAN dataset
  • Mel-spectrogram + CNN
  • FastAPI: audio upload → genre

Mini-loyiha 4: Time Series Anomaly Detection

  • Server metrics (CPU, RAM)
  • LSTM autoencoder
  • Real-time alert system

Quiz

Fundamentals

  1. Backpropagation qanday ishlaydi (chain rule)?
  2. Vanishing gradient nima va qanday hal qilinadi?
  3. Batch size va learning rate orasidagi munosabat?
  4. Why ReLU > Sigmoid (modern NN'larda)?
  5. Dropout test paytida nima qiladi?

PyTorch

  1. model.eval() va torch.no_grad() farqi?
  2. state_dict() nimani saqlaydi?
  3. DataLoader da num_workers va pin_memory ta'siri?
  4. Mixed precision (AMP) qachon foyda beradi?
  5. TorchScript va ONNX export'ning afzallik/kamchiligi?

CNN

  1. 3x3 kernel nima uchun keng tarqalgan?
  2. Max va Average pooling qachon qaysi birini ishlatasiz?
  3. ResNet'ning skip connection'i nima uchun ishlatiladi?
  4. EfficientNet'ning compound scaling'i nima?
  5. Receptive field nima va qanday hisoblanadi?

RNN

  1. RNN va Feedforward NN farqi?
  2. LSTM gate'lari va vazifalari?
  3. Bidirectional RNN qachon foyda beradi?
  4. Why gradient clipping is critical for RNN?
  5. RNN'dan Transformer'ga ko'chish sabablari?

✅ Oy 3 oxiri checklist

  • Pure NumPy bilan oddiy NN yozdim
  • PyTorch'da nn.Module va training loop
  • TensorFlow/Keras bilan tanishlik
  • CNN bilan image classification (CIFAR-10 yoki o'xshash)
  • Transfer learning (pretrained model bilan)
  • RNN/LSTM bilan sequence task (time series yoki text)
  • W&B yoki TensorBoard'da experiment tracking
  • FastAPI'da DL model serving (CPU yoki GPU'da)
  • Capstone loyiha GitHub'da
  • LinkedIn'ga post

Tabriklayman! Oy 4 — Computer Vision + NLP ga o'tamiz.