Skip to content

Latest commit

 

History

History
445 lines (329 loc) · 15.9 KB

README.md

File metadata and controls

445 lines (329 loc) · 15.9 KB
Raw

Chex

CI status docs pypi

Chex is a library of utilities for helping to write reliable JAX code.

This includes utils to help:

  • Instrument your code (e.g. assertions, warnings)
  • Debug (e.g. transforming pmaps in vmaps within a context manager).
  • Test JAX code across many variants (e.g. jitted vs non-jitted).

Installation

You can install the latest released version of Chex from PyPI via:

pip install chex

or you can install the latest development version from GitHub:

pip install git+https://github.com/deepmind/chex.git

Modules Overview

Dataclass (dataclass.py)

Dataclasses are a popular construct introduced by Python 3.7 to allow to easily specify typed data structures with minimal boilerplate code. They are not, however, compatible with JAX and dm-tree out of the box.

In Chex we provide a JAX-friendly dataclass implementation reusing python dataclasses.

Chex implementation of dataclass registers dataclasses as internal PyTree nodes to ensure compatibility with JAX data structures.

In addition, we provide a class wrapper that exposes dataclasses as collections.Mapping descendants which allows to process them (e.g. (un-)flatten) in dm-tree methods as usual Python dictionaries. See @mappable_dataclass docstring for more details.

Example:

@chex.dataclass
class Parameters:
  x: chex.ArrayDevice
  y: chex.ArrayDevice

parameters = Parameters(
    x=jnp.ones((2, 2)),
    y=jnp.ones((1, 2)),
)

# Dataclasses can be treated as JAX pytrees
jax.tree_util.tree_map(lambda x: 2.0 * x, parameters)

# and as mappings by dm-tree
tree.flatten(parameters)

NOTE: Unlike standard Python 3.7 dataclasses, Chex dataclasses cannot be constructed using positional arguments. They support construction arguments provided in the same format as the Python dict constructor. Dataclasses can be converted to tuples with the from_tuple and to_tuple methods if necessary.

parameters = Parameters(
    jnp.ones((2, 2)),
    jnp.ones((1, 2)),
)
# ValueError: Mappable dataclass constructor doesn't support positional args.

Assertions (asserts.py)

One limitation of PyType annotations for JAX is that they do not support the specification of DeviceArray ranks, shapes or dtypes. Chex includes a number of functions that allow flexible and concise specification of these properties.

E.g. suppose you want to ensure that all tensors t1, t2, t3 have the same shape, and that tensors t4, t5 have rank 2 and (3 or 4), respectively.

chex.assert_equal_shape([t1, t2, t3])
chex.assert_rank([t4, t5], [2, {3, 4}])

More examples:

from chex import assert_shape, assert_rank, ...

assert_shape(x, (2, 3))                # x has shape (2, 3)
assert_shape([x, y], [(), (2,3)])      # x is scalar and y has shape (2, 3)

assert_rank(x, 0)                      # x is scalar
assert_rank([x, y], [0, 2])            # x is scalar and y is a rank-2 array
assert_rank([x, y], {0, 2})            # x and y are scalar OR rank-2 arrays

assert_type(x, int)                    # x has type `int` (x can be an array)
assert_type([x, y], [int, float])      # x has type `int` and y has type `float`

assert_equal_shape([x, y, z])          # x, y, and z have equal shapes

assert_trees_all_close(tree_x, tree_y) # values and structure of trees match
assert_tree_all_finite(tree_x)         # all tree_x leaves are finite

assert_devices_available(2, 'gpu')     # 2 GPUs available
assert_tpu_available()                 # at least 1 TPU available

assert_numerical_grads(f, (x, y), j)   # f^{(j)}(x, y) matches numerical grads

See asserts.py documentation to find all supported assertions.

If you cannot find a specific assertion, please consider making a pull request or openning an issue on the bug tracker.

Optional Arguments

All chex assertions support the following optional kwargs for manipulating the emitted exception messages:

  • custom_message: A string to include into the emitted exception messages.
  • include_default_message: Whether to include the default Chex message into the emitted exception messages.
  • exception_type: An exception type to use. AssertionError by default.

For example, the following code:

dataset = load_dataset()
params = init_params()
for i in range(num_steps):
  params = update_params(params, dataset.sample())
  chex.assert_tree_all_finite(params,
                              custom_message=f'Failed at iteration {i}.',
                              exception_type=ValueError)

will raise a ValueError that includes a step number when params get polluted with NaNs or Nones.

Static and Value (aka Runtime) Assertions

Chex divides all assertions into 2 classes: static and value assertions.

  1. static assertions use anything except concrete values of tensors. Examples: assert_shape, assert_trees_all_equal_dtypes, assert_max_traces.

  2. value assertions require access to tensor values, which are not available during JAX tracing (see HowJAX primitives work), thus such assertion need special treatment in a jitted code.

To enable value assertions in a jitted function, it can be decorated with chex.chexify() wrapper. Example:

  @chex.chexify
  @jax.jit
  def logp1_abs_safe(x: chex.Array) -> chex.Array:
    chex.assert_tree_all_finite(x)
    return jnp.log(jnp.abs(x) + 1)

  logp1_abs_safe(jnp.ones(2))  # OK
  logp1_abs_safe(jnp.array([jnp.nan, 3]))  # FAILS (in async mode)

  # The error will be raised either at the next line OR at the next
  # `logp1_abs_safe` call. See the docs for more detain on async mode.
  logp1_abs_safe.wait_checks()  # Wait for the (async) computation to complete.

See this docstring for more detail on chex.chexify().

JAX Tracing Assertions

JAX re-traces JIT'ted function every time the structure of passed arguments changes. Often this behavior is inadvertent and leads to a significant performance drop which is hard to debug. @chex.assert_max_traces decorator asserts that the function is not re-traced more than n times during program execution.

Global trace counter can be cleared by calling chex.clear_trace_counter(). This function be used to isolate unittests relying on @chex.assert_max_traces.

Examples:

  @jax.jit
  @chex.assert_max_traces(n=1)
  def fn_sum_jitted(x, y):
    return x + y

  fn_sum_jitted(jnp.zeros(3), jnp.zeros(3))  # tracing for the 1st time - OK
  fn_sum_jitted(jnp.zeros([6, 7]), jnp.zeros([6, 7]))  # AssertionError!

Can be used with jax.pmap() as well:

  def fn_sub(x, y):
    return x - y

  fn_sub_pmapped = jax.pmap(chex.assert_max_traces(fn_sub, n=10))

See HowJAX primitives work section for more information about tracing.

Warnings (warnigns.py)

In addition to hard assertions Chex also offers utilities to add common warnings, such as specific types of deprecation warnings.

Test variants (variants.py)

JAX relies extensively on code transformation and compilation, meaning that it can be hard to ensure that code is properly tested. For instance, just testing a python function using JAX code will not cover the actual code path that is executed when jitted, and that path will also differ whether the code is jitted for CPU, GPU, or TPU. This has been a source of obscure and hard to catch bugs where XLA changes would lead to undesirable behaviours that however only manifest in one specific code transformation.

Variants make it easy to ensure that unit tests cover different ‘variations’ of a function, by providing a simple decorator that can be used to repeat any test under all (or a subset) of the relevant code transformations.

E.g. suppose you want to test the output of a function fn with or without jit. You can use chex.variants to run the test with both the jitted and non-jitted version of the function by simply decorating a test method with @chex.variants, and then using self.variant(fn) in place of fn in the body of the test.

def fn(x, y):
  return x + y
...

class ExampleTest(chex.TestCase):

  @chex.variants(with_jit=True, without_jit=True)
  def test(self):
    var_fn = self.variant(fn)
    self.assertEqual(fn(1, 2), 3)
    self.assertEqual(var_fn(1, 2), fn(1, 2))

If you define the function in the test method, you may also use self.variant as a decorator in the function definition. For example:

class ExampleTest(chex.TestCase):

  @chex.variants(with_jit=True, without_jit=True)
  def test(self):
    @self.variant
    def var_fn(x, y):
       return x + y

    self.assertEqual(var_fn(1, 2), 3)

Example of parameterized test:

from absl.testing import parameterized

# Could also be:
#  `class ExampleParameterizedTest(chex.TestCase, parameterized.TestCase):`
#  `class ExampleParameterizedTest(chex.TestCase):`
class ExampleParameterizedTest(parameterized.TestCase):

  @chex.variants(with_jit=True, without_jit=True)
  @parameterized.named_parameters(
      ('case_positive', 1, 2, 3),
      ('case_negative', -1, -2, -3),
  )
  def test(self, arg_1, arg_2, expected):
    @self.variant
    def var_fn(x, y):
       return x + y

    self.assertEqual(var_fn(arg_1, arg_2), expected)

Chex currently supports the following variants:

  • with_jit -- applies jax.jit() transformation to the function.
  • without_jit -- uses the function as is, i.e. identity transformation.
  • with_device -- places all arguments (except specified in ignore_argnums argument) into device memory before applying the function.
  • without_device -- places all arguments in RAM before applying the function.
  • with_pmap -- applies jax.pmap() transformation to the function (see notes below).

See documentation in variants.py for more details on the supported variants. More examples can be found in variants_test.py.

Variants notes

  • Test classes that use @chex.variants must inherit from chex.TestCase (or any other base class that unrolls tests generators within TestCase, e.g. absl.testing.parameterized.TestCase).

  • [jax.vmap] All variants can be applied to a vmapped function; please see an example in variants_test.py (test_vmapped_fn_named_params and test_pmap_vmapped_fn).

  • [@chex.all_variants] You can get all supported variants by using the decorator @chex.all_variants.

  • [with_pmap variant] jax.pmap(fn) (doc) performs parallel map of fn onto multiple devices. Since most tests run in a single-device environment (i.e. having access to a single CPU or GPU), in which case jax.pmap is a functional equivalent to jax.jit, with_pmap variant is skipped by default (although it works fine with a single device). Below we describe a way to properly test fn if it is supposed to be used in multi-device environments (TPUs or multiple CPUs/GPUs). To disable skipping with_pmap variants in case of a single device, add --chex_skip_pmap_variant_if_single_device=false to your test command.

Fakes (fake.py)

Debugging in JAX is made more difficult by code transformations such as jit and pmap, which introduce optimizations that make code hard to inspect and trace. It can also be difficult to disable those transformations during debugging as they can be called at several places in the underlying code. Chex provides tools to globally replace jax.jit with a no-op transformation and jax.pmap with a (non-parallel) jax.vmap, in order to more easily debug code in a single-device context.

For example, you can use Chex to fake pmap and have it replaced with a vmap. This can be achieved by wrapping your code with a context manager:

with chex.fake_pmap():
  @jax.pmap
  def fn(inputs):
    ...

  # Function will be vmapped over inputs
  fn(inputs)

The same functionality can also be invoked with start and stop:

fake_pmap = chex.fake_pmap()
fake_pmap.start()
... your jax code ...
fake_pmap.stop()

In addition, you can fake a real multi-device test environment with a multi-threaded CPU. See section Faking multi-device test environments for more details.

See documentation in fake.py and examples in fake_test.py for more details.

Faking multi-device test environments

In situations where you do not have easy access to multiple devices, you can still test parallel computation using single-device multi-threading.

In particular, one can force XLA to use a single CPU's threads as separate devices, i.e. to fake a real multi-device environment with a multi-threaded one. These two options are theoretically equivalent from XLA perspective because they expose the same interface and use identical abstractions.

Chex has a flag chex_n_cpu_devices that specifies a number of CPU threads to use as XLA devices.

To set up a multi-threaded XLA environment for absl tests, define setUpModule function in your test module:

def setUpModule():
  chex.set_n_cpu_devices()

Now you can launch your test with python test.py --chex_n_cpu_devices=N to run it in multi-device regime. Note that all tests within a module will have an access to N devices.

More examples can be found in variants_test.py, fake_test.py and fake_set_n_cpu_devices_test.py.

Using named dimension sizes.

Chex comes with a small utility that allows you to package a collection of dimension sizes into a single object. The basic idea is:

dims = chex.Dimensions(B=batch_size, T=sequence_len, E=embedding_dim)
...
chex.assert_shape(arr, dims['BTE'])

String lookups are translated integer tuples. For instance, let's say batch_size == 3, sequence_len = 5 and embedding_dim = 7, then

dims['BTE'] == (3, 5, 7)
dims['B'] == (3,)
dims['TTBEE'] == (5, 5, 3, 7, 7)
...

You can also assign dimension sizes dynamically as follows:

dims['XY'] = some_matrix.shape
dims.Z = 13

For more examples, see chex.Dimensions documentation.

Citing Chex

This repository is part of the DeepMind JAX Ecosystem, to cite Chex please use the DeepMind JAX Ecosystem citation.