Skillbooks / Library / 02
Machine learning engineering
Model Refactoring for Software Engineers
Take a model that already works and make it smaller, faster and deployable without breaking it.
Every chapter treats a model change the way you would treat any other engineering change: define the problem, inspect the implementation, make one controlled modification, test the result, and keep a rollback path. Worked examples are in Python and PyTorch. The methods apply across vision, language, audio, and image and video diffusion.
- Language
- English
- Length
- Seventeen chapters and four appendices, around 16,000 words
- Contains
- Worked examples, runnable companion code, the recorded evidence behind the results, a technical glossary, and fifty primary sources
- Formats
- Reading edition in the browser, PDF, Markdown, and a companion archive with code and evidence
- Price
- Free. No account, no payment
Editions
Read it in the browser, or reproduce itRead online
Reading edition
The whole handbook in the browser, with chapter navigation down the side. Follows light or dark, and returns you to where you stopped reading.
Download
The complete handbook, typeset and ready to read or print. Nothing to sign up for. The file is yours.
Plain text
Markdown
The whole manuscript as plain text, with internal links throughout. Search it, quote it, diff it, or adapt it for your own team.
Reproduce the results
Companion archive
Everything in one file: the manuscript, the PDF, the three lab scripts, every recorded result and checkpoint, and the checksums to verify them.
Contents
Seventeen chapters, four appendicesCode and evidence
Open any file directlyThe book does not ask you to take its results on trust. The scripts that produced them are here, along with the raw output they wrote, so you can read the code, compare it against the recorded numbers, and run it yourself.
What the companion code runs
The included experiments are a trained synthetic classifier and small mechanism tests for convolution, attention and chunking, with the original recorded outputs, raw timing samples and generated checkpoints. They demonstrate the method end to end on something you can actually run on a laptop.
They are not a production benchmark. No GPU, Android or NPU measurement is included, and the export and quantization templates are given as recipes rather than executed runs. Adapt them to your own model and validate against your own evaluation suite, which is the discipline the book argues for throughout.