Sphinx is apparently quite picky about consistent use of headers: you can't
skip a header level. We were getting warnings like "WARNING: Title level
inconsistent" in the docs build, and sub-headers weren't showing up on this
page after the first section.
* Use a whitelist to restrict visibility in top-level jax namespace.
The goal of this change is to capture the way the world is (i.e., not break users), and separately we will work on fixing users to avoid accidentally-exported APIs.
* Remove usage of xla_client.{Computation,ComputationBuilder}.
ComputationBuilder is a fairly pointless wrapper class that mimics an outdated version of the the C++ XLA API. It dates back from when we used to have SWIG bindings and needed to write a non-trivial Python shim to keep the interface pleasant to use. Now that we have pybind11-based bindings that are reasonably ergonomic by themselves, we don't need the wrapper class. Instead, we can simply call the pybind11-wrapped C++ API directly, removing the impedance mismatch between the C++ and Python APIs and allowing us to delete the Python ComputationBuilder class.
Similarly we can delete xla_client.Computation for the same reasons; it doesn't do anything useful on top of the C++ API.
* Fix minor typo in cell
One of the arguments to `hvp` wasn't being used, which made the example slightly confusing.
* Fix both definitions of hvp in the autodiff cookbook.
Co-authored-by: Peter Hawkins <phawkins@google.com>
Introduced two new constructors for PartialVal: unknown and known.
These should make it easier to read the code where we construct
PartialVal:
* instead of PartialVal((aval, core.unit) we use PartialVal.unknown(aval)
* instead of PartialVal((None, pval)) we use PartialVal.known(pval)
Also disabled some new tests in random_tests.py on Mac. They segfault,
apparently due to the same issue #432.
* Minor update to docs; trigger readthedocs
* Updated Common Gotchas notebook
Handle errors explicitly, otherwise it is too hard to test the notebook by 'Run all'
* Added a section about pure functions to Common Gotchas
Long, long ago, when JAX was first born, we realized that we couldn't
transpose this jaxpr:
{ lambda ; a.
let b = reduce_sum[ axes=(0,) ] a
in b }
The problem was that the transpose of a reduce-sum is a broadcast, but
because jaxprs didn't have shape information available, we didn't know
what input shape to broadcast to!
Our hack was to have the primitives that required shape information for
transposition to acquire it into their parameters, so that we'd produce
jaxprs like this one:
{ lambda ; a.
let b = reduce_sum[ axes=(0,)
input_shape=(3,) ] a
in b }
That's not only aesthetically unpleasant, but also it meant we were
limiting an (unused) capability of the system: ideally we should be able
to trace a reduce-sum jaxpr without specializing on shape information
(e.g. at the Unshaped level) and only require shape specialization for
transposition. (Good thing no one actually traces at Unshaped...)
But at long last @chr1sj0nes in #2299 added avals to jaxprs, so that
shape information (or whatever information with which the jaxpr was
specialized out of Python) is in the jaxpr itself. So we could finally
remove these shapes-in-params warts!
That's exactly what this commit does!
Co-authored-by: Roy Frostig <frostig@google.com>
Co-authored-by: Roy Frostig <frostig@google.com>
