Tools & Libraries: Start Here
A short index: which package does which job and where it is taught. NumPy, pandas, plotting, SQL and scikit-learn stay in their own topics; this section adds PyTorch, Hugging Face and gradient boosting.
The ML engineer's toolbox
This page only points to where things are taught; it does not repeat them. A handful of packages do almost all everyday ML work. You do not need to memorise their APIs: you need to know which tool fits which job, the core pattern of each, and how to read its documentation.
| Job | Package | Taught in |
|---|---|---|
| Arrays and maths | NumPy | NumPy: Fast Arrays |
| Tables: load, clean, join, group | pandas | pandas |
| Charts | Matplotlib, seaborn | Plotting |
| Data from databases | SQL (SQLite, PostgreSQL…) | SQL for ML |
| Classical ML: preprocessing, models, evaluation | scikit-learn | scikit-learn Essentials |
| Best models for table data | XGBoost, LightGBM (and scikit-learn's HistGradientBoosting) | Gradient Boosting Libraries |
| Deep learning | PyTorch | PyTorch Essentials, then Deep Learning in Practice and the Deep Learning track |
| Pretrained models: text, vision, audio, LLMs | Hugging Face (transformers, datasets, peft) | Hugging Face, then the NLP track |
| Code, versions, environments, tests | terminal, git, venv/conda, pytest | Your Developer Toolkit |
What about TensorFlow/Keras and JAX? PyTorch is now the most common framework in research and in most job adverts. Keras (now multi-backend) and JAX are used in some companies and labs; once you know PyTorch, their concepts (tensors, layers, autograd, optimisers, training loops) transfer directly, so learn them if a job needs them.
How to learn any new library quickly
- Find the core pattern: scikit-learn is
fit / predict; PyTorch isModule + DataLoader + loop; Hugging Face isfrom_pretrained. Most of the API follows from it. - Run the official quickstart end to end before reading anything else.
- Read the docstring of the functions you use:
help(f)orf?in Jupyter. Check the defaults: many bugs are surprising defaults. - Pin versions in
requirements.txt; APIs change (for example, parameter names in Hugging Face and XGBoost have changed between versions). - Check what the library did: print shapes, look at a few outputs, and compare with a tiny case you can compute by hand.
Quiz
1. Which package would you reach for first to train a model on a table of customer records?
2. You need a pretrained text classifier quickly. Which library?
3. Why pin package versions in a project?
📚 Further reading
Trusted, mostly free sources: official documentation, well-known textbooks and courses. When you want to confirm or go deeper on anything here, start with these.
- scikit-learn User Guide ↗ Official documentation for every model
- PyTorch: Learn the Basics (official tutorial) ↗ tensors, datasets, models, autograd, optimisation, saving
- Hugging Face Transformers documentation ↗ pipelines, Auto classes, Trainer
- The Missing Semester of Your CS Education (MIT) ↗ Shell, git, editors and tools