How-to guides

A how-to guide gets a competent reader from a goal they already have to a result. It assumes they know what they are trying to do and answers only how.

What belongs here

A page belongs in this section when its title could begin with "How to" and a reader would arrive already wanting the outcome.

• It starts from a real goal, stated in the reader's terms rather than the implementation's.
• It gives a sequence that works, and it is honest about the choices along the way — unlike a tutorial, a how-to may say "if your problem is sparse, do this instead," because its reader can judge.
• It stops when the goal is met. Background that does not change what the reader types belongs in an explanation.
• It links to reference for the parameters it mentions rather than restating them, so a default has exactly one home.

What does not belong here

• Teaching the subject from nothing. That is a tutorial. A how-to may assume its reader knows what a learning rate is.
• Why the method works. That is an explanation. A guide to running a gradient check says what tolerance to compare against; the derivation of the finite-difference approximation belongs elsewhere.
• The full parameter list of a function. That is reference, and it is generated from the docstrings, so a guide that copies it will go stale.
• A page organised around a component rather than a goal. "The Adam optimizer" is a reference or explanation page. "How to choose an optimizer" is a how-to.

Contents

• How to choose an optimizer — from the practical guidelines and optimizer selection sections of the optimization algorithms document.
• How to run a gradient check — from the gradient checking procedure, including how to read the resulting relative error.
• How to tune a learning rate — from the learning rate finder material and the tuning starting strategy.
• How to run a hyperparameter search — from the search strategy guidance, covering search space design and the pitfalls documented alongside it.
• How to choose a gated cell — picking between the LSTM, the GRU, and the minimal gated recurrences when a gate is what the task needs.
• How to choose a sequence model — choosing between a recurrent network, a state-space model, and attention, from the sequence's length and what it costs to process.
• How to train a language model and generate text — the training loop, what perplexity tells you, and how to sample from the trained distribution.
• How to train word vectors on your own corpus — from tokens to trained, inspected vectors with skip-gram or GloVe, and the knobs that decide whether they come out good.
• How to choose an embedding model — from your task, corpus, language, and budget to an embedding family, the consumption contract to pin alongside it, and the measurement that decides whether to keep it.