Repository navigation
Module 05: Advanced Topics
高级专题 - 深入 Keras 高级 API，掌握模型优化与工程化最佳实践。
| Attribute | Value |
|---|---|
| Prerequisites | Modules 02-04 |
| Duration | 2-3 weeks |
| Notebooks | 12+ |
| Difficulty | ⭐⭐⭐⭐ Advanced |
| Topic | Key Concepts |
|---|---|
| Multi-Input Models | 多输入融合 |
| Multi-Output Models | 多任务学习 |
| Shared Layers | 权重共享 |
| Residual Connections | 残差连接 |
| Callback | Purpose |
|---|---|
ModelCheckpoint |
保存最佳模型 |
EarlyStopping |
早停防过拟合 |
ReduceLROnPlateau |
动态学习率 |
TensorBoard |
训练可视化 |
LearningRateScheduler |
学习率调度 |
| Technique | Tool |
|---|---|
| Hyperparameter Tuning | Optuna, Ray Tune |
| Pruning | TF Model Optimization |
| Quantization | TFLite, ONNX |
| Knowledge Distillation | Teacher-Student |
生成式模型 - 探索生成式模型的原理与应用。
- Latent Space Learning
- Reparameterization Trick
- KL Divergence Loss
- DCGAN: Deep Convolutional GAN
- WGAN: Wasserstein GAN
- StyleGAN: Style-based Generation
- DDPM: Denoising Diffusion
- Score Matching
- Stable Diffusion
强化学习 - 掌握强化学习核心算法。
| Category | Algorithms |
|---|---|
| Value-Based | Q-Learning, DQN, Double DQN, Dueling DQN |
| Policy-Based | REINFORCE, PPO, SAC |
| Actor-Critic | A2C, A3C, TD3 |
| Model-Based | World Models, MuZero |
理论笔记 - 全面的理论参考资料。
- ReLU, LeakyReLU, PReLU, ELU
- GELU, Swish, Mish, SiLU
- Sigmoid, Tanh, Softmax
- Cross-Entropy, Focal Loss
- MSE, MAE, Huber Loss
- Contrastive Loss, Triplet Loss
- CNN Design Patterns
- Transformer Variants
- Efficient Architectures
实战项目 - 通过真实项目检验学习成果。
| Category | Projects |
|---|---|
| ML Basics | Titanic, Otto Classification |
| Computer Vision | MNIST CNN, Image Classification |
| NLP | Sentiment Analysis, NER, Translation |
| Time Series | Temperature, Stock Prediction |
| Kaggle | Gold Medal Solutions |
