🧱 Version 1: Practical ML Engineer Track

Phase 0: Training Stack Ownership

• [ ] Build a training pipeline from scratch (dataset → model → training loop → logging)
• [ ] Train a small CNN or MLP end-to-end with TensorBoard or equivalent logs

Phase 1: Architectural Breadth

• [ ] Complete D2L chapters 5–10 and replicate each example in PyTorch
• [ ] Modify 5+ Keras example models (e.g., change width, activations, add dropout/batchnorm)
• [ ] Write 3-paragraph reports on each modified Keras model's performance changes

Phase 2: Improvement Skills

• [ ] Take 3 real Kaggle datasets and:
  • [ ] Start from weak architecture
  • [ ] Iteratively improve it through architectural changes
• [ ] For each dataset, deliver:
  • [ ] Before/after training curves
  • [ ] Architecture diagrams
  • [ ] Explanation of changes and why they helped

Phase 3: Modern Patterns

• [ ] Re-implement ResNet18 from scratch (not using torchvision.models)
• [ ] Clone a modern architecture (e.g., ConvNeXt, EfficientNet, MobileNet) from its paper/code
• [ ] Experiment: Add LayerNorm or residuals to a basic CNN and analyze effect
• [ ] Write 1 short experimental paper: “What happens when I add Transformer-style normalization to a CNN?”

Phase 4: Model Surgery

• [ ] Deliberately break a model (e.g., bad initialization, missing normalization)
• [ ] Diagnose failure through training logs