Documentation

Dive into Deep Learning · §1.7

Using an unfamiliar API
discover · inspect · read · verify.

A procedure for consulting an API

Motivation

No book covers an entire framework, and libraries change across releases. Four steps answer most routine questions from within a notebook: discover available names, inspect the interface, read the documentation, and verify the behavior with a small example.

Use the official reference and tutorial pages for the supported interface.

Figuring out an unfamiliar API: a repeatable loopDiscoverdir() · TabInspecthelp() · ?Readdocs · source ??Verifyrun a quick teststill stuck? refine the query and loop

dir discovers what exists

Discover

Know roughly where a tool should live, but not its name? dir lists a module’s contents; the names alone sketch what is on offer:

pprint([name for name in dir(tf.random)
        if not name.startswith('_')][:20], compact=True)
['Algorithm', 'Generator', 'all_candidate_sampler', 'categorical',
 'create_rng_state', 'experimental', 'fixed_unigram_candidate_sampler',
 'fold_in', 'gamma', 'get_global_generator',
 'learned_unigram_candidate_sampler', 'log_uniform_candidate_sampler', 'normal',
 'poisson', 'set_global_generator', 'set_seed', 'shuffle', 'split',
 'stateless_binomial', 'stateless_categorical']

Skip the _-prefixed internals. In a notebook, module. + Tab gives the same list, filtered as you type, usually the fastest way to turn up a name.

help reads the signature; ?? reads the source

Inspect · read

help(...) prints the docstring: arguments, defaults, return value, often an example.

help(tf.ones)

In Jupyter, ones? opens the docstring in a side pane, and ones?? shows the source code, which can clarify a terse or ambiguous docstring.

A tiny run settles it

Verify

Docstrings can drift out of date; verify the current behavior with a small call:

tf.ones(4)
<tf.Tensor: shape=(4,), dtype=float32, numpy=array([1., 1., 1., 1.], dtype=float32)>

The result has the documented shape and values. The discover → inspect → read → verify loop remains useful as APIs change.

Coding assistants enter the same loop

Assistants

An assistant may produce a plausible function and call. Treat the suggestion as a candidate to check before building on it.

Glance at the signature (help / ?), then run a small example. A suggestion that survives both is one you can rely on.

Recap

Wrap-up

  • Discover with dir (or Tab-completion).
  • Inspect with help / ?; read the source with ??.
  • Verify with a tiny run.
  • Assistant answers enter the same loop before you rely on them.