venv/lib/python3.11/site-packages/chex/_src/dataclass_test.py

# Copyright 2020 DeepMind Technologies Limited. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Tests for `dataclass.py`."""

# pytype: disable=wrong-keyword-args  # dataclass_transform

import copy
import dataclasses
import pickle
import sys
from typing import Any, Generic, Mapping, TypeVar

from absl.testing import absltest
from absl.testing import parameterized
from chex._src import asserts
from chex._src import dataclass
from chex._src import pytypes
import cloudpickle
import jax
import numpy as np
import tree

chex_dataclass = dataclass.dataclass
mappable_dataclass = dataclass.mappable_dataclass
orig_dataclass = dataclasses.dataclass


@chex_dataclass
class NestedDataclass():
  c: pytypes.ArrayDevice
  d: pytypes.ArrayDevice


@chex_dataclass
class PostInitDataclass:
  a: pytypes.ArrayDevice

  def __post_init__(self):
    if not self.a > 0:
      raise ValueError('a should be > than 0')


@chex_dataclass
class ReverseOrderNestedDataclass():
  # The order of c and d are switched comapred to NestedDataclass.
  d: pytypes.ArrayDevice
  c: pytypes.ArrayDevice


@chex_dataclass
class Dataclass():
  a: NestedDataclass
  b: pytypes.ArrayDevice


@chex_dataclass(frozen=True)
class FrozenDataclass():
  a: NestedDataclass
  b: pytypes.ArrayDevice


def dummy_dataclass(factor=1., frozen=False):
  class_ctor = FrozenDataclass if frozen else Dataclass
  return class_ctor(
      a=NestedDataclass(
          c=factor * np.ones((3,), dtype=np.float32),
          d=factor * np.ones((4,), dtype=np.float32)),
      b=factor * 2 * np.ones((5,), dtype=np.float32))


def _dataclass_instance_fields(dcls_instance):
  """Serialization-friendly version of dataclasses.fields for instances."""
  attribute_dict = dcls_instance.__dict__
  fields = []
  for field in dcls_instance.__dataclass_fields__.values():
    if field.name in attribute_dict:  # Filter pseudo-fields.
      fields.append(field)
  return fields


@orig_dataclass
class ClassWithoutMap:
  k: dict  # pylint:disable=g-bare-generic

  def some_method(self, *args):
    raise RuntimeError('ClassWithoutMap.some_method() was called.')


def _get_mappable_dataclasses(test_type):
  """Generates shallow and nested mappable dataclasses."""

  class Class:
    """Shallow class."""

    k_tuple: tuple  # pylint:disable=g-bare-generic
    k_dict: dict  # pylint:disable=g-bare-generic

    def some_method(self, *args):
      raise RuntimeError('Class.some_method() was called.')

  class NestedClass:
    """Nested class."""

    k_any: Any
    k_int: int
    k_str: str
    k_arr: np.ndarray
    k_dclass_with_map: Class
    k_dclass_no_map: ClassWithoutMap
    k_dict_factory: dict = dataclasses.field(  # pylint:disable=g-bare-generic,invalid-field-call
        default_factory=lambda: dict(x='x', y='y'))
    k_default: str = 'default_str'
    k_non_init: int = dataclasses.field(default=1, init=False)  # pylint:disable=g-bare-generic,invalid-field-call
    k_init_only: dataclasses.InitVar[int] = 10

    def some_method(self, *args):
      raise RuntimeError('NestedClassWithMap.some_method() was called.')

    def __post_init__(self, k_init_only):
      self.k_non_init = self.k_int * k_init_only

  if test_type == 'chex':
    cls = chex_dataclass(Class, mappable_dataclass=True)
    nested_cls = chex_dataclass(NestedClass, mappable_dataclass=True)
  elif test_type == 'original':
    cls = mappable_dataclass(orig_dataclass(Class))
    nested_cls = mappable_dataclass(orig_dataclass(NestedClass))
  else:
    raise ValueError(f'Unknown test type: {test_type}')

  return cls, nested_cls


@parameterized.named_parameters(('_original', 'original'), ('_chex', 'chex'))
class MappableDataclassTest(parameterized.TestCase):

  def _init_testdata(self, test_type):
    """Initializes test data."""
    map_cls, nested_map_cls = _get_mappable_dataclasses(test_type)

    self.dcls_with_map_inner = map_cls(
        k_tuple=(1, 2), k_dict=dict(k1=32, k2=33))
    self.dcls_with_map_inner_inc = map_cls(
        k_tuple=(2, 3), k_dict=dict(k1=33, k2=34))

    self.dcls_no_map = ClassWithoutMap(k=dict(t='t', t2='t2'))
    self.dcls_with_map = nested_map_cls(
        k_any=None,
        k_int=1,
        k_str='test_str',
        k_arr=np.array(16),
        k_dclass_with_map=self.dcls_with_map_inner,
        k_dclass_no_map=self.dcls_no_map)

    self.dcls_with_map_inc_ints = nested_map_cls(
        k_any=None,
        k_int=2,
        k_str='test_str',
        k_arr=np.array(16),
        k_dclass_with_map=self.dcls_with_map_inner_inc,
        k_dclass_no_map=self.dcls_no_map,
        k_default='default_str')

    self.dcls_flattened_with_path = [
        (('k_any',), None),
        (('k_arr',), np.array(16)),
        (('k_dclass_no_map',), self.dcls_no_map),
        (('k_dclass_with_map', 'k_dict', 'k1'), 32),
        (('k_dclass_with_map', 'k_dict', 'k2'), 33),
        (('k_dclass_with_map', 'k_tuple', 0), 1),
        (('k_dclass_with_map', 'k_tuple', 1), 2),
        (('k_default',), 'default_str'),
        (('k_dict_factory', 'x'), 'x'),
        (('k_dict_factory', 'y'), 'y'),
        (('k_int',), 1),
        (('k_non_init',), 10),
        (('k_str',), 'test_str'),
    ]

    self.dcls_flattened_with_path_up_to = [
        (('k_any',), None),
        (('k_arr',), np.array(16)),
        (('k_dclass_no_map',), self.dcls_no_map),
        (('k_dclass_with_map',), self.dcls_with_map_inner),
        (('k_default',), 'default_str'),
        (('k_dict_factory', 'x'), 'x'),
        (('k_dict_factory', 'y'), 'y'),
        (('k_int',), 1),
        (('k_non_init',), 10),
        (('k_str',), 'test_str'),
    ]

    self.dcls_flattened = [v for (_, v) in self.dcls_flattened_with_path]
    self.dcls_flattened_up_to = [
        v for (_, v) in self.dcls_flattened_with_path_up_to
    ]
    self.dcls_tree_size = 18
    self.dcls_tree_size_no_dicts = 14

  def testFlattenAndUnflatten(self, test_type):
    self._init_testdata(test_type)

    self.assertEqual(self.dcls_flattened, tree.flatten(self.dcls_with_map))
    self.assertEqual(
        self.dcls_with_map,
        tree.unflatten_as(self.dcls_with_map_inc_ints, self.dcls_flattened))

    dataclass_in_seq = [34, self.dcls_with_map, [1, 2]]
    dataclass_in_seq_flat = [34] + self.dcls_flattened + [1, 2]
    self.assertEqual(dataclass_in_seq_flat, tree.flatten(dataclass_in_seq))
    self.assertEqual(dataclass_in_seq,
                     tree.unflatten_as(dataclass_in_seq, dataclass_in_seq_flat))

  def testFlattenUpTo(self, test_type):
    self._init_testdata(test_type)
    structure = copy.copy(self.dcls_with_map)
    structure.k_dclass_with_map = None  # Do not flatten 'k_dclass_with_map'
    self.assertEqual(self.dcls_flattened_up_to,
                     tree.flatten_up_to(structure, self.dcls_with_map))

  def testFlattenWithPath(self, test_type):
    self._init_testdata(test_type)

    self.assertEqual(
        tree.flatten_with_path(self.dcls_with_map),
        self.dcls_flattened_with_path)

  def testFlattenWithPathUpTo(self, test_type):
    self._init_testdata(test_type)
    structure = copy.copy(self.dcls_with_map)
    structure.k_dclass_with_map = None  # Do not flatten 'k_dclass_with_map'
    self.assertEqual(
        tree.flatten_with_path_up_to(structure, self.dcls_with_map),
        self.dcls_flattened_with_path_up_to)

  def testMapStructure(self, test_type):
    self._init_testdata(test_type)

    add_one_to_ints_fn = lambda x: x + 1 if isinstance(x, int) else x
    mapped_inc_ints = tree.map_structure(add_one_to_ints_fn, self.dcls_with_map)

    self.assertEqual(self.dcls_with_map_inc_ints, mapped_inc_ints)
    self.assertEqual(self.dcls_with_map_inc_ints.k_non_init,
                     self.dcls_with_map_inc_ints.k_int * 10)
    self.assertEqual(mapped_inc_ints.k_non_init, mapped_inc_ints.k_int * 10)

  def testMapStructureUpTo(self, test_type):
    self._init_testdata(test_type)

    structure = copy.copy(self.dcls_with_map)
    structure.k_dclass_with_map = None  # Do not map over 'k_dclass_with_map'
    add_one_to_ints_fn = lambda x: x + 1 if isinstance(x, int) else x
    mapped_inc_ints = tree.map_structure_up_to(structure, add_one_to_ints_fn,
                                               self.dcls_with_map)

    # k_dclass_with_map should be passed through unchanged
    class_with_map = self.dcls_with_map.k_dclass_with_map
    self.dcls_with_map_inc_ints.k_dclass_with_map = class_with_map
    self.assertEqual(self.dcls_with_map_inc_ints, mapped_inc_ints)
    self.assertEqual(self.dcls_with_map_inc_ints.k_non_init,
                     self.dcls_with_map_inc_ints.k_int * 10)
    self.assertEqual(mapped_inc_ints.k_non_init, mapped_inc_ints.k_int * 10)

  def testMapStructureWithPath(self, test_type):
    self._init_testdata(test_type)

    add_one_to_ints_fn = lambda path, x: x + 1 if isinstance(x, int) else x
    mapped_inc_ints = tree.map_structure_with_path(add_one_to_ints_fn,
                                                   self.dcls_with_map)

    self.assertEqual(self.dcls_with_map_inc_ints, mapped_inc_ints)
    self.assertEqual(self.dcls_with_map_inc_ints.k_non_init,
                     self.dcls_with_map_inc_ints.k_int * 10)
    self.assertEqual(mapped_inc_ints.k_non_init, mapped_inc_ints.k_int * 10)

  def testMapStructureWithPathUpTo(self, test_type):
    self._init_testdata(test_type)

    structure = copy.copy(self.dcls_with_map)
    structure.k_dclass_with_map = None  # Do not map over 'k_dclass_with_map'
    add_one_to_ints_fn = lambda path, x: x + 1 if isinstance(x, int) else x
    mapped_inc_ints = tree.map_structure_with_path_up_to(
        structure, add_one_to_ints_fn, self.dcls_with_map)

    # k_dclass_with_map should be passed through unchanged
    class_with_map = self.dcls_with_map.k_dclass_with_map
    self.dcls_with_map_inc_ints.k_dclass_with_map = class_with_map

    self.assertEqual(self.dcls_with_map_inc_ints, mapped_inc_ints)
    self.assertEqual(self.dcls_with_map_inc_ints.k_non_init,
                     self.dcls_with_map_inc_ints.k_int * 10)
    self.assertEqual(mapped_inc_ints.k_non_init, mapped_inc_ints.k_int * 10)

  def testTraverse(self, test_type):
    self._init_testdata(test_type)

    visited = []
    tree.traverse(visited.append, self.dcls_with_map, top_down=False)
    self.assertLen(visited, self.dcls_tree_size)

    visited_without_dicts = []

    def visit_without_dicts(x):
      visited_without_dicts.append(x)
      return 'X' if isinstance(x, dict) else None

    tree.traverse(visit_without_dicts, self.dcls_with_map, top_down=True)
    self.assertLen(visited_without_dicts, self.dcls_tree_size_no_dicts)

  def testIsDataclass(self, test_type):
    self._init_testdata(test_type)

    self.assertTrue(dataclasses.is_dataclass(self.dcls_no_map))
    self.assertTrue(dataclasses.is_dataclass(self.dcls_with_map))
    self.assertTrue(
        dataclasses.is_dataclass(self.dcls_with_map.k_dclass_with_map))
    self.assertTrue(
        dataclasses.is_dataclass(self.dcls_with_map.k_dclass_no_map))


class DataclassesTest(parameterized.TestCase):

  @parameterized.parameters([True, False])
  def test_dataclass_tree_leaves(self, frozen):
    obj = dummy_dataclass(frozen=frozen)
    self.assertLen(jax.tree_util.tree_leaves(obj), 3)

  @parameterized.parameters([True, False])
  def test_dataclass_tree_map(self, frozen):
    factor = 5.
    obj = dummy_dataclass(frozen=frozen)
    target_obj = dummy_dataclass(factor=factor, frozen=frozen)
    asserts.assert_trees_all_close(
        jax.tree_util.tree_map(lambda t: factor * t, obj), target_obj)

  def test_tree_flatten_with_keys(self):
    obj = dummy_dataclass()
    keys_and_leaves, treedef = jax.tree_util.tree_flatten_with_path(obj)
    self.assertEqual(
        [k for k, _ in keys_and_leaves],
        [
            (jax.tree_util.GetAttrKey('a'), jax.tree_util.GetAttrKey('c')),
            (jax.tree_util.GetAttrKey('a'), jax.tree_util.GetAttrKey('d')),
            (jax.tree_util.GetAttrKey('b'),),
        ],
    )
    leaves = [l for _, l in keys_and_leaves]
    new_obj = treedef.unflatten(leaves)
    self.assertEqual(new_obj, obj)

  def test_tree_map_with_keys(self):
    obj = dummy_dataclass()
    key_value_list, unused_treedef = jax.tree_util.tree_flatten_with_path(obj)
    # Convert a list of key-value tuples to a dict.
    flat_obj = dict(key_value_list)

    def f(path, x):
      value = flat_obj[path]
      np.testing.assert_allclose(value, x)
      return path

    out = jax.tree_util.tree_map_with_path(f, obj)
    self.assertEqual(
        out.a.c, (jax.tree_util.GetAttrKey('a'), jax.tree_util.GetAttrKey('c'))
    )
    self.assertEqual(
        out.a.d, (jax.tree_util.GetAttrKey('a'), jax.tree_util.GetAttrKey('d'))
    )
    self.assertEqual(out.b, (jax.tree_util.GetAttrKey('b'),))

  def test_tree_map_with_keys_traversal_order(self):
    # pytype: disable=wrong-arg-types
    obj = ReverseOrderNestedDataclass(d=1, c=2)
    # pytype: enable=wrong-arg-types
    leaves = []
    def f(_, x):
      leaves.append(x)

    jax.tree_util.tree_map_with_path(f, obj)
    self.assertEqual(leaves, jax.tree_util.tree_leaves(obj))

  @parameterized.parameters([True, False])
  def test_dataclass_replace(self, frozen):
    factor = 5.
    obj = dummy_dataclass(frozen=frozen)
    # pytype: disable=attribute-error  # dataclass_transform
    obj = obj.replace(a=obj.a.replace(c=factor * obj.a.c))
    obj = obj.replace(a=obj.a.replace(d=factor * obj.a.d))
    obj = obj.replace(b=factor * obj.b)
    target_obj = dummy_dataclass(factor=factor, frozen=frozen)
    asserts.assert_trees_all_close(obj, target_obj)
    # pytype: enable=attribute-error

  def test_dataclass_requires_kwargs_by_default(self):
    factor = 1.0
    with self.assertRaisesRegex(
        ValueError,
        "Mappable dataclass constructor doesn't support positional args.",
    ):
      Dataclass(
          NestedDataclass(
              c=factor * np.ones((3,), dtype=np.float32),
              d=factor * np.ones((4,), dtype=np.float32),
          ),
          factor * 2 * np.ones((5,), dtype=np.float32),
      )

  def test_dataclass_mappable_dataclass_false(self):
    factor = 1.0

    @chex_dataclass(mappable_dataclass=False)
    class NonMappableDataclass:
      a: NestedDataclass
      b: pytypes.ArrayDevice

    NonMappableDataclass(
        NestedDataclass(
            c=factor * np.ones((3,), dtype=np.float32),
            d=factor * np.ones((4,), dtype=np.float32),
        ),
        factor * 2 * np.ones((5,), dtype=np.float32),
    )

  def test_inheritance_is_possible_thanks_to_kw_only(self):
    if sys.version_info.minor < 10:  # Feature only available for Python >= 3.10
      return

    @chex_dataclass(kw_only=True)
    class Base:
      default: int = 1

    @chex_dataclass(kw_only=True)
    class Child(Base):
      non_default: int

    Child(non_default=2)

  def test_unfrozen_dataclass_is_mutable(self):
    factor = 5.
    obj = dummy_dataclass(frozen=False)
    obj.a.c = factor * obj.a.c
    obj.a.d = factor * obj.a.d
    obj.b = factor * obj.b
    target_obj = dummy_dataclass(factor=factor, frozen=False)
    asserts.assert_trees_all_close(obj, target_obj)

  def test_frozen_dataclass_raise_error(self):
    factor = 5.
    obj = dummy_dataclass(frozen=True)
    obj.a.c = factor * obj.a.c  # mutable since obj.a is not frozen.
    with self.assertRaisesRegex(dataclass.FrozenInstanceError,
                                'cannot assign to field'):
      obj.b = factor * obj.b  # raises error because obj is frozen.

  @parameterized.named_parameters(
      ('frozen', True),
      ('mutable', False),
  )
  def test_get_and_set_state(self, frozen):

    @chex_dataclass(frozen=frozen)
    class SimpleClass():
      data: int = 1

    obj_a = SimpleClass(data=1)
    state = getattr(obj_a, '__getstate__')()
    obj_b = SimpleClass(data=2)
    getattr(obj_b, '__setstate__')(state)
    self.assertEqual(obj_a, obj_b)

  def test_unexpected_kwargs(self):

    @chex_dataclass()
    class SimpleDataclass:
      a: int
      b: int = 2

    SimpleDataclass(a=1, b=3)
    with self.assertRaisesRegex(ValueError, 'init.*got unexpected kwargs'):
      SimpleDataclass(a=1, b=3, c=4)  # pytype: disable=wrong-keyword-args

  def test_tuple_conversion(self):

    @chex_dataclass()
    class SimpleDataclass:
      b: int
      a: int

    obj = SimpleDataclass(a=2, b=1)
    self.assertSequenceEqual(getattr(obj, 'to_tuple')(), (1, 2))

    obj2 = getattr(SimpleDataclass, 'from_tuple')((1, 2))
    self.assertEqual(obj.a, obj2.a)
    self.assertEqual(obj.b, obj2.b)

  @parameterized.named_parameters(
      ('frozen', True),
      ('mutable', False),
  )
  def test_tuple_rev_conversion(self, frozen):
    obj = dummy_dataclass(frozen=frozen)
    asserts.assert_trees_all_close(
        type(obj).from_tuple(obj.to_tuple()),  # pytype: disable=attribute-error
        obj,
    )

  @parameterized.named_parameters(
      ('frozen', True),
      ('mutable', False),
  )
  def test_inheritance(self, frozen):

    @chex_dataclass(frozen=frozen)
    class Base:
      x: int

    @chex_dataclass(frozen=frozen)
    class Derived(Base):
      y: int

    base_obj = Base(x=1)
    self.assertNotIsInstance(base_obj, Derived)
    self.assertIsInstance(base_obj, Base)

    derived_obj = Derived(x=1, y=2)
    self.assertIsInstance(derived_obj, Derived)
    self.assertIsInstance(derived_obj, Base)

  def test_inheritance_from_empty_frozen_base(self):

    @chex_dataclass(frozen=True)
    class FrozenBase:
      pass

    @chex_dataclass(frozen=True)
    class DerivedFrozen(FrozenBase):
      j: int

    df = DerivedFrozen(j=2)
    self.assertIsInstance(df, FrozenBase)

    with self.assertRaisesRegex(
        TypeError, 'cannot inherit non-frozen dataclass from a frozen one'):

      # pylint:disable=unused-variable
      @chex_dataclass
      class DerivedMutable(FrozenBase):
        j: int

      # pylint:enable=unused-variable

  def test_disallowed_fields(self):
    # pylint:disable=unused-variable
    with self.assertRaisesRegex(ValueError, 'dataclass fields are disallowed'):

      @chex_dataclass(mappable_dataclass=False)
      class InvalidNonMappable:
        from_tuple: int

    @chex_dataclass(mappable_dataclass=False)
    class ValidMappable:
      get: int

    with self.assertRaisesRegex(ValueError, 'dataclass fields are disallowed'):

      @chex_dataclass(mappable_dataclass=True)
      class InvalidMappable:
        get: int
        from_tuple: int

    # pylint:enable=unused-variable

  @parameterized.parameters(True, False)
  def test_flatten_is_leaf(self, is_mappable):

    @chex_dataclass(mappable_dataclass=is_mappable)
    class _InnerDcls:
      v_1: int
      v_2: int

    @chex_dataclass(mappable_dataclass=is_mappable)
    class _Dcls:
      str_val: str
      # pytype: disable=invalid-annotation  # enable-bare-annotations
      inner_dcls: _InnerDcls
      dct: Mapping[str, _InnerDcls]
      # pytype: enable=invalid-annotation  # enable-bare-annotations

    dcls = _Dcls(
        str_val='test',
        inner_dcls=_InnerDcls(v_1=1, v_2=11),
        dct={
            'md1': _InnerDcls(v_1=2, v_2=22),
            'md2': _InnerDcls(v_1=3, v_2=33)
        })

    def _is_leaf(value) -> bool:
      # Must not traverse over integers.
      self.assertNotIsInstance(value, int)
      return isinstance(value, (_InnerDcls, str))

    leaves = jax.tree_util.tree_flatten(dcls, is_leaf=_is_leaf)[0]
    self.assertCountEqual(
        (dcls.str_val, dcls.inner_dcls, dcls.dct['md1'], dcls.dct['md2']),
        leaves)

    asserts.assert_trees_all_equal_structs(
        jax.tree_util.tree_map(lambda x: x, dcls, is_leaf=_is_leaf), dcls)

  def test_decorator_alias(self):
    # Make sure, that creating a decorator alias works correctly.
    configclass = chex_dataclass(frozen=True)

    @configclass
    class Foo:
      bar: int = 1
      toto: int = 2

    @configclass
    class Bar:
      bar: int = 1
      toto: int = 2

    # Verify that both Foo and Bar are correctly registered with jax.tree_util.
    self.assertLen(jax.tree_util.tree_flatten(Foo())[0], 2)
    self.assertLen(jax.tree_util.tree_flatten(Bar())[0], 2)

  @parameterized.named_parameters(
      ('mappable', True),
      ('not_mappable', False),
  )
  def test_generic_dataclass(self, mappable):
    T = TypeVar('T')

    @chex_dataclass(mappable_dataclass=mappable)
    class GenericDataclass(Generic[T]):
      a: T  # pytype: disable=invalid-annotation  # enable-bare-annotations

    obj = GenericDataclass(a=np.array([1.0, 1.0]))
    asserts.assert_trees_all_close(obj.a, 1.0)

  def test_mappable_eq_override(self):

    @chex_dataclass(mappable_dataclass=True)
    class EqDataclass:
      a: pytypes.ArrayDevice

      def __eq__(self, other):
        if isinstance(other, EqDataclass):
          return other.a[0] == self.a[0]
        return False

    obj1 = EqDataclass(a=np.array([1.0, 1.0]))
    obj2 = EqDataclass(a=np.array([1.0, 0.0]))
    obj3 = EqDataclass(a=np.array([0.0, 1.0]))
    self.assertEqual(obj1, obj2)
    self.assertNotEqual(obj1, obj3)

  @parameterized.parameters([NestedDataclass, ReverseOrderNestedDataclass])
  def test_dataclass_instance_fields(self, dcls):
    obj = dcls(c=1, d=2)
    self.assertSequenceEqual(
        dataclasses.fields(obj), _dataclass_instance_fields(obj))

  @parameterized.parameters((pickle, NestedDataclass),
                            (cloudpickle, ReverseOrderNestedDataclass))
  def test_roundtrip_serialization(self, serialization_lib, dcls):
    obj = dcls(c=1, d=2)
    obj_fields = [
        (f.name, getattr(obj, f.name)) for f in dataclasses.fields(obj)
    ]
    self.assertLen(obj_fields, 2)
    obj2 = serialization_lib.loads(serialization_lib.dumps(obj))
    obj2_fields = [(f.name, getattr(obj2, f.name))
                   for f in _dataclass_instance_fields(obj2)]
    self.assertSequenceEqual(obj_fields, obj2_fields)
    self.assertSequenceEqual(jax.tree_util.tree_leaves(obj2), [1, 2])

    obj3 = jax.tree_util.tree_map(lambda x: x, obj2)
    obj3_fields = [(f.name, getattr(obj3, f.name))
                   for f in _dataclass_instance_fields(obj3)]
    self.assertSequenceEqual(obj_fields, obj3_fields)
    self.assertSequenceEqual(jax.tree_util.tree_leaves(obj3), [1, 2])

  @parameterized.parameters([NestedDataclass, ReverseOrderNestedDataclass])
  def test_flatten_roundtrip_ordering(self, dcls):
    obj = dcls(c=1, d=2)
    leaves, treedef = jax.tree_util.tree_flatten(obj)
    self.assertSequenceEqual(leaves, [1, 2])
    obj2 = jax.tree_util.tree_unflatten(treedef, leaves)
    self.assertSequenceEqual(dataclasses.fields(obj2), dataclasses.fields(obj))

  def test_flatten_respects_post_init(self):
    obj = PostInitDataclass(a=1)  # pytype: disable=wrong-arg-types
    with self.assertRaises(ValueError):
      _ = jax.tree_util.tree_map(lambda x: 0, obj)

  @parameterized.parameters([False, True])
  def test_keys_and_values_type(self, frozen):
    obj = dummy_dataclass(frozen=frozen)
    self.assertEqual(
        type(obj.keys()),  # pytype: disable=attribute-error
        type({}.keys()),
    )
    self.assertEqual(
        type(obj.values()),  # pytype: disable=attribute-error
        type({}.values()),
    )

  @parameterized.parameters([False, True])
  def test_keys_and_values_override(self, frozen):
    @chex_dataclass(frozen=frozen)
    class _Dataclass:
      x: int
      values: int

    obj = _Dataclass(x=1, values=2)
    self.assertEqual(
        list(obj.keys()),  # pytype: disable=attribute-error
        ['x', 'values'],
    )
    self.assertEqual(obj.values, 2)


if __name__ == '__main__':
  absltest.main()