Certificate

# ===== Skills checklist =====
# After the PyTorch track you should be able to:
# [x] Load + batch data with Dataset + DataLoader
# [x] Build models with nn.Module
# [x] Write a training loop with mixed precision + gradient clipping
# [x] Use lr schedulers + early stopping properly
# [x] Save + load checkpoints (model + optimizer + scheduler + scaler)
# [x] Train across multiple GPUs with DistributedDataParallel
# [x] Profile with torch.profiler
# [x] Export to TorchScript and ONNX
# [x] Move models to mobile via PyTorch Mobile / ExecuTorch
# [x] Use Lightning for boilerplate-free training (optional)
# [x] Inspect gradients + activations during training

# ===== Bookmark =====
# - https://pytorch.org/docs                official docs
# - https://pytorch.org/tutorials           official tutorials
# - https://lightning.ai                    PyTorch Lightning
# - https://github.com/karpathy/nanoGPT     minimal GPT in PyTorch
# - https://github.com/huggingface/transformers
# - 'Deep Learning with PyTorch' (Stevens, Antiga, Viehmann)

# ===== Portfolio project (8-16 hours) =====
# Pick a realistic problem and ship it end-to-end:
# 1) Dataset + dataloader (real data, not toy)
# 2) Model class + a reproducible training script
# 3) Reproducibility seed + a 'how to retrain' command in README
# 4) Mixed precision + gradient clipping
# 5) Best-checkpoint saving
# 6) Held-out test set + reported metric with a confidence interval
# 7) Inference API (FastAPI) that loads the artifact
# 8) Small demo UI (Gradio / Streamlit) on Hugging Face Spaces
# 9) ONNX export for cross-framework consumption

# Bonus:
# - Multi-GPU training script via torchrun
# - SHAP / Captum explanations for individual predictions
# - A blog post documenting design decisions

# ===== What 'good' looks like =====
# - Train + test metric reported HONESTLY (test set held out)
# - Reproducible: one command from raw data to trained model
# - Inference latency reported (e.g. p95 30ms on CPU)
# - README explains hyperparams + why
# - .gitignore excludes /data/, /checkpoints/, .ipynb_checkpoints

# ===== Common mistakes to avoid =====
# - .item() inside training loop (silent perf killer)
# - Forgetting model.eval() before inference
# - Mixed precision without GradScaler
# - DataParallel (use DistributedDataParallel)
# - Not exporting the model (just shipping a notebook)
# - Reporting validation as if it were test

# ===== Next steps =====
# - Hugging Face Transformers for NLP / vision SOTA
# - PyTorch Lightning to remove training-loop boilerplate
# - DeepSpeed / FSDP for large-scale training
# - torch.compile (2.0+) for faster training
# - Pytorch Mobile / ExecuTorch for on-device inference

# ===== Self-test =====
# If you can:
# 1) Diagnose a slow training run via torch.profiler
# 2) Resume training from a checkpoint that includes scaler + scheduler state
# 3) Export a trained model to ONNX and run it from Node.js
# you have completed the track. Ship the portfolio piece and call it done.

# ===== Track wrap-up =====
# PyTorch is the production ML framework most teams ship with in 2026. The
# habits that matter:
# - Reproducibility: seeds, fixed environment, retrain-from-scratch script
# - Mixed precision + grad clipping + scheduler from the start
# - Profile before optimising
# - Export to a portable format (ONNX, TorchScript, TFLite)
# Get those right and you ship models, not notebooks.

# /certificate/pytorch