Oven logo

Oven

optree0.13.1

Published

Optimized PyTree Utilities.

pip install optree

Package Downloads

Weekly DownloadsMonthly Downloads

Requires Python

>=3.7

Dependencies

OpTree

Python 3.7+ PyPI GitHub Workflow Status GitHub Workflow Status Codecov Documentation Status Downloads GitHub Repo Stars

Optimized PyTree Utilities.

Table of Contents

Installation

Install from PyPI (PyPI / Status):

pip3 install --upgrade optree

Install from conda-forge (conda-forge):

conda install -c conda-forge optree

Install the latest version from GitHub:

pip3 install git+https://github.com/metaopt/optree.git#egg=optree

Or, clone this repo and install manually:

git clone --depth=1 https://github.com/metaopt/optree.git
cd optree
pip3 install .

Compiling from the source requires Python 3.7+, a compiler (gcc / clang / icc / cl.exe) that supports C++20 and a cmake installation.

PyTrees

A PyTree is a recursive structure that can be an arbitrarily nested Python container (e.g., tuple, list, dict, OrderedDict, NamedTuple, etc.) or an opaque Python object. The key concepts of tree operations are tree flattening and its inverse (tree unflattening). Additional tree operations can be performed based on these two basic functions (e.g., tree_map = tree_unflatten ∘ map ∘ tree_flatten).

Tree flattening is traversing the entire tree in a left-to-right depth-first manner and returning the leaves of the tree in a deterministic order.

>>> tree = {'b': (2, [3, 4]), 'a': 1, 'c': 5, 'd': 6}
>>> optree.tree_flatten(tree)
([1, 2, 3, 4, 5, 6], PyTreeSpec({'a': *, 'b': (*, [*, *]), 'c': *, 'd': *}))
>>> optree.tree_flatten(1)
([1], PyTreeSpec(*))
>>> optree.tree_flatten(None)
([], PyTreeSpec(None))

This usually implies that the equal pytrees return equal lists of leaves and the same tree structure. See also section Key Ordering for Dictionaries.

>>> {'a': [1, 2], 'b': [3]} == {'b': [3], 'a': [1, 2]}
True
>>> optree.tree_leaves({'a': [1, 2], 'b': [3]}) == optree.tree_leaves({'b': [3], 'a': [1, 2]})
True
>>> optree.tree_structure({'a': [1, 2], 'b': [3]}) == optree.tree_structure({'b': [3], 'a': [1, 2]})
True

Tree Nodes and Leaves

A tree is a collection of non-leaf nodes and leaf nodes, where the leaf nodes have no children to flatten. optree.tree_flatten(...) will flatten the tree and return a list of leaf nodes while the non-leaf nodes will store in the tree specification.

Built-in PyTree Node Types

OpTree out-of-box supports the following Python container types in the registry:

which are considered non-leaf nodes in the tree. Python objects that the type is not registered will be treated as leaf nodes. The registry lookup uses the is operator to determine whether the type is matched. So subclasses will need to explicitly register in the registry, otherwise, an object of that type will be considered a leaf. The NoneType is a special case discussed in section None is non-leaf Node vs. None is Leaf.

Registering a Container-like Custom Type as Non-leaf Nodes

A container-like Python type can be registered in the type registry with a pair of functions that specify:

  • flatten_func(container) -> (children, metadata, entries): convert an instance of the container type to a (children, metadata, entries) triple, where children is an iterable of subtrees and entries is an iterable of path entries of the container (e.g., indices or keys).
  • unflatten_func(metadata, children) -> container: convert such a pair back to an instance of the container type.

The metadata is some necessary data apart from the children to reconstruct the container, e.g., the keys of the dictionary (the children are values).

The entries can be omitted (only returns a pair) or is optional to implement (returns None). If so, use range(len(children)) (i.e., flat indices) as path entries of the current node. The function signature can be flatten_func(container) -> (children, metadata) or flatten_func(container) -> (children, metadata, None).

The following examples show how to register custom types and utilize them for tree_flatten and tree_map. Please refer to section Notes about the PyTree Type Registry for more information.

# Registry a Python type with lambda functions
optree.register_pytree_node(
    set,
    # (set) -> (children, metadata, None)
    lambda s: (sorted(s), None, None),
    # (metadata, children) -> (set)
    lambda _, children: set(children),
    namespace='set',
)

# Register a Python type into a namespace
import torch

class Torch2NumpyEntry(optree.PyTreeEntry):
    def __call__(self, obj):
        assert self.entry == 0
        return obj.cpu().detach().numpy()

    def codify(self, node=''):
        assert self.entry == 0
        return f'{node}.cpu().detach().numpy()'

optree.register_pytree_node(
    torch.Tensor,
    # (tensor) -> (children, metadata)
    flatten_func=lambda tensor: (
        (tensor.cpu().detach().numpy(),),
        {'dtype': tensor.dtype, 'device': tensor.device, 'requires_grad': tensor.requires_grad},
    ),
    # (metadata, children) -> tensor
    unflatten_func=lambda metadata, children: torch.tensor(children[0], **metadata),
    path_entry_type=Torch2NumpyEntry,
    namespace='torch2numpy',
)

>>> tree = {'weight': torch.ones(size=(1, 2)).cuda(), 'bias': torch.zeros(size=(2,))}
>>> tree
{'weight': tensor([[1., 1.]], device='cuda:0'), 'bias': tensor([0., 0.])}

# Flatten without specifying the namespace
>>> optree.tree_flatten(tree)  # `torch.Tensor`s are leaf nodes
([tensor([0., 0.]), tensor([[1., 1.]], device='cuda:0')], PyTreeSpec({'bias': *, 'weight': *}))

# Flatten with the namespace
>>> leaves, treespec = optree.tree_flatten(tree, namespace='torch2numpy')
>>> leaves, treespec
(
    [array([0., 0.], dtype=float32), array([[1., 1.]], dtype=float32)],
    PyTreeSpec(
        {
            'bias': CustomTreeNode(Tensor[{'dtype': torch.float32, 'device': device(type='cpu'), 'requires_grad': False}], [*]),
            'weight': CustomTreeNode(Tensor[{'dtype': torch.float32, 'device': device(type='cuda', index=0), 'requires_grad': False}], [*])
        },
        namespace='torch2numpy'
    )
)

# `entries` are not defined and use `range(len(children))`
>>> optree.tree_paths(tree, namespace='torch2numpy')
[('bias', 0), ('weight', 0)]

# Custom path entry type defines the pytree access behavior
>>> optree.tree_accessors(tree, namespace='torch2numpy')
[
    PyTreeAccessor(*['bias'].cpu().detach().numpy(), (MappingEntry(key='bias', type=<class 'dict'>), Torch2NumpyEntry(entry=0, type=<class 'torch.Tensor'>))),
    PyTreeAccessor(*['weight'].cpu().detach().numpy(), (MappingEntry(key='weight', type=<class 'dict'>), Torch2NumpyEntry(entry=0, type=<class 'torch.Tensor'>)))
]

# Unflatten back to a copy of the original object
>>> optree.tree_unflatten(treespec, leaves)
{'weight': tensor([[1., 1.]], device='cuda:0'), 'bias': tensor([0., 0.])}

Users can also extend the pytree registry by decorating the custom class and defining an instance method tree_flatten and a class method tree_unflatten.

from collections import UserDict

@optree.register_pytree_node_class(namespace='mydict')
class MyDict(UserDict):
    TREE_PATH_ENTRY_TYPE = optree.MappingEntry  # used by accessor APIs

    def tree_flatten(self):  # -> (children, metadata, entries)
        reversed_keys = sorted(self.keys(), reverse=True)
        return (
            [self[key] for key in reversed_keys],  # children
            reversed_keys,  # metadata
            reversed_keys,  # entries
        )

    @classmethod
    def tree_unflatten(cls, metadata, children):
        return cls(zip(metadata, children))

>>> tree = MyDict(b=4, a=(2, 3), c=MyDict({'d': 5, 'f': 6}))

# Flatten without specifying the namespace
>>> optree.tree_flatten_with_path(tree)  # `MyDict`s are leaf nodes
(
    [()],
    [MyDict(b=4, a=(2, 3), c=MyDict({'d': 5, 'f': 6}))],
    PyTreeSpec(*)
)

# Flatten with the namespace
>>> optree.tree_flatten_with_path(tree, namespace='mydict')
(
    [('c', 'f'), ('c', 'd'), ('b',), ('a', 0), ('a', 1)],
    [6, 5, 4, 2, 3],
    PyTreeSpec(
        CustomTreeNode(MyDict[['c', 'b', 'a']], [CustomTreeNode(MyDict[['f', 'd']], [*, *]), *, (*, *)]),
        namespace='mydict'
    )
)
>>> optree.tree_flatten_with_accessor(tree, namespace='mydict')
(
    [
        PyTreeAccessor(*['c']['f'], (MappingEntry(key='c', type=<class 'MyDict'>), MappingEntry(key='f', type=<class 'MyDict'>))),
        PyTreeAccessor(*['c']['d'], (MappingEntry(key='c', type=<class 'MyDict'>), MappingEntry(key='d', type=<class 'MyDict'>))),
        PyTreeAccessor(*['b'], (MappingEntry(key='b', type=<class 'MyDict'>),)),
        PyTreeAccessor(*['a'][0], (MappingEntry(key='a', type=<class 'MyDict'>), SequenceEntry(index=0, type=<class 'tuple'>))),
        PyTreeAccessor(*['a'][1], (MappingEntry(key='a', type=<class 'MyDict'>), SequenceEntry(index=1, type=<class 'tuple'>)))
    ],
    [6, 5, 4, 2, 3],
    PyTreeSpec(
        CustomTreeNode(MyDict[['c', 'b', 'a']], [CustomTreeNode(MyDict[['f', 'd']], [*, *]), *, (*, *)]),
        namespace='mydict'
    )
)

Notes about the PyTree Type Registry

There are several key attributes of the pytree type registry:

  1. The type registry is per-interpreter-dependent. This means registering a custom type in the registry affects all modules that use OpTree.

[!WARNING] For safety reasons, a namespace must be specified while registering a custom type. It is used to isolate the behavior of flattening and unflattening a pytree node type. This is to prevent accidental collisions between different libraries that may register the same type.

  1. The elements in the type registry are immutable. Users can neither register the same type twice in the same namespace (i.e., update the type registry), nor remove a type from the type registry. To update the behavior of an already registered type, simply register it again with another namespace.

  2. Users cannot modify the behavior of already registered built-in types listed in Built-in PyTree Node Types, such as key order sorting for dict and collections.defaultdict.

  3. Inherited subclasses are not implicitly registered. The registry lookup uses type(obj) is registered_type rather than isinstance(obj, registered_type). Users need to register the subclasses explicitly. To register all subclasses, it is easy to implement with metaclass or __init_subclass__, for example:

    from collections import UserDict

@optree.register_pytree_node_class(namespace='mydict')
class MyDict(UserDict):
    TREE_PATH_ENTRY_TYPE = optree.MappingEntry  # used by accessor APIs

    def __init_subclass__(cls):  # define this in the base class
        super().__init_subclass__()
        # Register a subclass to namespace 'mydict'
        optree.register_pytree_node_class(cls, namespace='mydict')

    def tree_flatten(self):  # -> (children, metadata, entries)
        reversed_keys = sorted(self.keys(), reverse=True)
        return (
            [self[key] for key in reversed_keys],  # children
            reversed_keys,  # metadata
            reversed_keys,  # entries
        )

    @classmethod
    def tree_unflatten(cls, metadata, children):
        return cls(zip(metadata, children))

# Subclasses will be automatically registered in namespace 'mydict'
class MyAnotherDict(MyDict):
    pass

    >>> tree = MyDict(b=4, a=(2, 3), c=MyAnotherDict({'d': 5, 'f': 6}))
>>> optree.tree_flatten_with_path(tree, namespace='mydict')
(
    [('c', 'f'), ('c', 'd'), ('b',), ('a', 0), ('a', 1)],
    [6, 5, 4, 2, 3],
    PyTreeSpec(
        CustomTreeNode(MyDict[['c', 'b', 'a']], [CustomTreeNode(MyAnotherDict[['f', 'd']], [*, *]), *, (*, *)]),
        namespace='mydict'
    )
)
>>> optree.tree_accessors(tree, namespace='mydict')
[
    PyTreeAccessor(*['c']['f'], (MappingEntry(key='c', type=<class 'MyDict'>), MappingEntry(key='f', type=<class 'MyAnotherDict'>))),
    PyTreeAccessor(*['c']['d'], (MappingEntry(key='c', type=<class 'MyDict'>), MappingEntry(key='d', type=<class 'MyAnotherDict'>))),
    PyTreeAccessor(*['b'], (MappingEntry(key='b', type=<class 'MyDict'>),)),
    PyTreeAccessor(*['a'][0], (MappingEntry(key='a', type=<class 'MyDict'>), SequenceEntry(index=0, type=<class 'tuple'>))),
    PyTreeAccessor(*['a'][1], (MappingEntry(key='a', type=<class 'MyDict'>), SequenceEntry(index=1, type=<class 'tuple'>)))
]

  4. Be careful about the potential infinite recursion of the custom flatten function. The returned children from the custom flatten function are considered subtrees. They will be further flattened recursively. The children can have the same type as the current node. Users must design their termination condition carefully.

    import numpy as np
import torch

optree.register_pytree_node(
    np.ndarray,
    # Children are nest lists of Python objects
    lambda array: (np.atleast_1d(array).tolist(), array.ndim == 0),
    lambda scalar, rows: np.asarray(rows) if not scalar else np.asarray(rows[0]),
    namespace='numpy1',
)

optree.register_pytree_node(
    np.ndarray,
    # Children are Python objects
    lambda array: (
        list(array.ravel()),  # list(1DArray[T]) -> List[T]
        dict(shape=array.shape, dtype=array.dtype)
    ),
    lambda metadata, children: np.asarray(children, dtype=metadata['dtype']).reshape(metadata['shape']),
    namespace='numpy2',
)

optree.register_pytree_node(
    np.ndarray,
    # Returns a list of `np.ndarray`s without termination condition
    lambda array: ([array.ravel()], array.dtype),
    lambda shape, children: children[0].reshape(shape),
    namespace='numpy3',
)

optree.register_pytree_node(
    torch.Tensor,
    # Children are nest lists of Python objects
    lambda tensor: (torch.atleast_1d(tensor).tolist(), tensor.ndim == 0),
    lambda scalar, rows: torch.tensor(rows) if not scalar else torch.tensor(rows[0])),
    namespace='torch1',
)

optree.register_pytree_node(
    torch.Tensor,
    # Returns a list of `torch.Tensor`s without termination condition
    lambda tensor: (
        list(tensor.view(-1)),  # list(1DTensor[T]) -> List[0DTensor[T]] (STILL TENSORS!)
        tensor.shape
    ),
    lambda shape, children: torch.stack(children).reshape(shape),
    namespace='torch2',
)

    >>> optree.tree_flatten(np.arange(9).reshape(3, 3), namespace='numpy1')
(
    [0, 1, 2, 3, 4, 5, 6, 7, 8],
    PyTreeSpec(
        CustomTreeNode(ndarray[False], [[*, *, *], [*, *, *], [*, *, *]]),
        namespace='numpy1'
    )
)
# Implicitly casts `float`s to `np.float64`
>>> optree.tree_map(lambda x: x + 1.5, np.arange(9).reshape(3, 3), namespace='numpy1')
array([[1.5, 2.5, 3.5],
       [4.5, 5.5, 6.5],
       [7.5, 8.5, 9.5]])

>>> optree.tree_flatten(np.arange(9).reshape(3, 3), namespace='numpy2')
(
    [0, 1, 2, 3, 4, 5, 6, 7, 8],
    PyTreeSpec(
        CustomTreeNode(ndarray[{'shape': (3, 3), 'dtype': dtype('int64')}], [*, *, *, *, *, *, *, *, *]),
        namespace='numpy2'
    )
)
# Explicitly casts `float`s to `np.int64`
>>> optree.tree_map(lambda x: x + 1.5, np.arange(9).reshape(3, 3), namespace='numpy2')
array([[1, 2, 3],
       [4, 5, 6],
       [7, 8, 9]])

# Children are also `np.ndarray`s, recurse without termination condition.
>>> optree.tree_flatten(np.arange(9).reshape(3, 3), namespace='numpy3')
Traceback (most recent call last):
    ...
RecursionError: Maximum recursion depth exceeded during flattening the tree.

>>> optree.tree_flatten(torch.arange(9).reshape(3, 3), namespace='torch1')
(
    [0, 1, 2, 3, 4, 5, 6, 7, 8],
    PyTreeSpec(
        CustomTreeNode(Tensor[False], [[*, *, *], [*, *, *], [*, *, *]]),
        namespace='torch1'
    )
)
# Implicitly casts `float`s to `torch.float32`
>>> optree.tree_map(lambda x: x + 1.5, torch.arange(9).reshape(3, 3), namespace='torch1')
tensor([[1.5000, 2.5000, 3.5000],
        [4.5000, 5.5000, 6.5000],
        [7.5000, 8.5000, 9.5000]])

# Children are also `torch.Tensor`s, recurse without termination condition.
>>> optree.tree_flatten(torch.arange(9).reshape(3, 3), namespace='torch2')
Traceback (most recent call last):
    ...
RecursionError: Maximum recursion depth exceeded during flattening the tree.

None is Non-leaf Node vs. None is Leaf

The None object is a special object in the Python language. It serves some of the same purposes as null (a pointer does not point to anything) in other programming languages, which denotes a variable is empty or marks default parameters. However, the None object is a singleton object rather than a pointer. It may also serve as a sentinel value. In addition, if a function has returned without any return value or the return statement is omitted, the function will also implicitly return the None object.

By default, the None object is considered a non-leaf node in the tree with arity 0, i.e., a non-leaf node that has no children. This is like the behavior of an empty tuple. While flattening a tree, it will remain in the tree structure definitions rather than in the leaves list.

>>> tree = {'b': (2, [3, 4]), 'a': 1, 'c': None, 'd': 5}
>>> optree.tree_flatten(tree)
([1, 2, 3, 4, 5], PyTreeSpec({'a': *, 'b': (*, [*, *]), 'c': None, 'd': *}))
>>> optree.tree_flatten(tree, none_is_leaf=True)
([1, 2, 3, 4, None, 5], PyTreeSpec({'a': *, 'b': (*, [*, *]), 'c': *, 'd': *}, NoneIsLeaf))
>>> optree.tree_flatten(1)
([1], PyTreeSpec(*))
>>> optree.tree_flatten(None)
([], PyTreeSpec(None))
>>> optree.tree_flatten(None, none_is_leaf=True)
([None], PyTreeSpec(*, NoneIsLeaf))

OpTree provides a keyword argument none_is_leaf to determine whether to consider the None object as a leaf, like other opaque objects. If none_is_leaf=True, the None object will be placed in the leaves list. Otherwise, the None object will remain in the tree specification (structure).

>>> import torch

>>> linear = torch.nn.Linear(in_features=3, out_features=2, bias=False)
>>> linear._parameters  # a container has None
OrderedDict({
    'weight': Parameter containing:
              tensor([[-0.6677,  0.5209,  0.3295],
                      [-0.4876, -0.3142,  0.1785]], requires_grad=True),
    'bias': None
})

>>> optree.tree_map(torch.zeros_like, linear._parameters)
OrderedDict({
    'weight': tensor([[0., 0., 0.],
                      [0., 0., 0.]]),
    'bias': None
})

>>> optree.tree_map(torch.zeros_like, linear._parameters, none_is_leaf=True)
Traceback (most recent call last):
    ...
TypeError: zeros_like(): argument 'input' (position 1) must be Tensor, not NoneType

>>> optree.tree_map(lambda t: torch.zeros_like(t) if t is not None else 0, linear._parameters, none_is_leaf=True)
OrderedDict({
    'weight': tensor([[0., 0., 0.],
                      [0., 0., 0.]]),
    'bias': 0
})

Key Ordering for Dictionaries

The built-in Python dictionary (i.e., builtins.dict) is an unordered mapping that holds the keys and values. The leaves of a dictionary are the values. Although since Python 3.6, the built-in dictionary is insertion ordered (PEP 468). The dictionary equality operator (==) does not check for key ordering. To ensure referential transparency that "equal dict" implies "equal ordering of leaves", the order of values of the dictionary is sorted by the keys. This behavior is also applied to collections.defaultdict.

>>> optree.tree_flatten({'a': [1, 2], 'b': [3]})
([1, 2, 3], PyTreeSpec({'a': [*, *], 'b': [*]}))
>>> optree.tree_flatten({'b': [3], 'a': [1, 2]})
([1, 2, 3], PyTreeSpec({'a': [*, *], 'b': [*]}))

If users want to keep the values in the insertion order in pytree traversal, they should use collections.OrderedDict, which will take the order of keys under consideration:

>>> OrderedDict([('a', [1, 2]), ('b', [3])]) == OrderedDict([('b', [3]), ('a', [1, 2])])
False
>>> optree.tree_flatten(OrderedDict([('a', [1, 2]), ('b', [3])]))
([1, 2, 3], PyTreeSpec(OrderedDict({'a': [*, *], 'b': [*]})))
>>> optree.tree_flatten(OrderedDict([('b', [3]), ('a', [1, 2])]))
([3, 1, 2], PyTreeSpec(OrderedDict({'b': [*], 'a': [*, *]})))

Since OpTree v0.9.0, the key order of the reconstructed output dictionaries from tree_unflatten is guaranteed to be consistent with the key order of the input dictionaries in tree_flatten.

>>> leaves, treespec = optree.tree_flatten({'b': [3], 'a': [1, 2]})
>>> leaves, treespec
([1, 2, 3], PyTreeSpec({'a': [*, *], 'b': [*]}))
>>> optree.tree_unflatten(treespec, leaves)
{'b': [3], 'a': [1, 2]}
>>> optree.tree_map(lambda x: x, {'b': [3], 'a': [1, 2]})
{'b': [3], 'a': [1, 2]}
>>> optree.tree_map(lambda x: x + 1, {'b': [3], 'a': [1, 2]})
{'b': [4], 'a': [2, 3]}

This property is also preserved during serialization/deserialization.

>>> leaves, treespec = optree.tree_flatten({'b': [3], 'a': [1, 2]})
>>> leaves, treespec
([1, 2, 3], PyTreeSpec({'a': [*, *], 'b': [*]}))
>>> restored_treespec = pickle.loads(pickle.dumps(treespec))
>>> optree.tree_unflatten(treespec, leaves)
{'b': [3], 'a': [1, 2]}
>>> optree.tree_unflatten(restored_treespec, leaves)
{'b': [3], 'a': [1, 2]}

[!NOTE] Note that there are no restrictions on the dict to require the keys to be comparable (sortable). There can be multiple types of keys in the dictionary. The keys are sorted in ascending order by key=lambda k: k first if capable otherwise fallback to key=lambda k: (f'{k.__class__.__module__}.{k.__class__.__qualname__}', k). This handles most cases.

>>> sorted({1: 2, 1.5: 1}.keys())
[1, 1.5]
>>> sorted({'a': 3, 1: 2, 1.5: 1}.keys())
Traceback (most recent call last):
    ...
TypeError: '<' not supported between instances of 'int' and 'str'
>>> sorted({'a': 3, 1: 2, 1.5: 1}.keys(), key=lambda k: (f'{k.__class__.__module__}.{k.__class__.__qualname__}', k))
[1.5, 1, 'a']

Benchmark

We benchmark the performance of:

  • tree flatten
  • tree unflatten
  • tree copy (i.e., unflatten(flatten(...)))
  • tree map

compared with the following libraries:

Average Time Cost (↓)OpTree (v0.9.0)JAX XLA (v0.4.6)PyTorch (v2.0.0)DM-Tree (v0.1.8)
Tree Flattenx1.002.3322.051.12
Tree UnFlattenx1.002.694.2816.23
Tree Flatten with Pathx1.0016.16Not Supported27.59
Tree Copyx1.002.569.9711.02
Tree Mapx1.002.569.5810.62
Tree Map (nargs)x1.002.89Not Supported31.33
Tree Map with Pathx1.007.23Not Supported19.66
Tree Map with Path (nargs)x1.006.56Not Supported29.61

All results are reported on a workstation with an AMD Ryzen 9 5950X CPU @ 4.45GHz in an isolated virtual environment with Python 3.10.9. Run with the following commands:

conda create --name optree-benchmark anaconda::python=3.10 --yes --no-default-packages
conda activate optree-benchmark
python3 -m pip install --editable '.[benchmark]' --extra-index-url https://download.pytorch.org/whl/cpu
python3 benchmark.py --number=10000 --repeat=5

The test inputs are nested containers (i.e., pytrees) extracted from torch.nn.Module objects. They are:

tiny_mlp = nn.Sequential(
    nn.Linear(1, 1, bias=True),
    nn.BatchNorm1d(1, affine=True, track_running_stats=True),
    nn.ReLU(),
    nn.Linear(1, 1, bias=False),
    nn.Sigmoid(),
)

and AlexNet, ResNet18, ResNet34, ResNet50, ResNet101, ResNet152, VisionTransformerH14 (ViT-H/14), and SwinTransformerB (Swin-B) from torchvsion. Please refer to benchmark.py for more details.

Tree Flatten

ModuleNodesOpTree (μs)JAX XLA (μs)PyTorch (μs)DM-Tree (μs)Speedup (J / O)Speedup (P / O)Speedup (D / O)
TinyMLP5329.7071.06583.6631.322.3919.651.05
AlexNet188103.92262.562304.36119.612.5322.171.15
ResNet18698368.06852.698440.31420.432.3222.931.14
ResNet341242644.961461.5514498.81712.812.2722.481.11
ResNet501702919.952080.5820995.961006.422.2622.821.09
ResNet10133171806.363996.9040314.121955.482.2122.321.08
ResNet15249322656.925812.3857775.532826.922.1921.751.06
ViT-H/1434201863.504418.2441334.642128.712.3722.181.14
Swin-B28811631.063944.1336131.542032.772.4222.151.25
Average2.3322.051.12

Tree UnFlatten

ModuleNodesOpTree (μs)JAX XLA (μs)PyTorch (μs)DM-Tree (μs)Speedup (J / O)Speedup (P / O)Speedup (D / O)
TinyMLP5355.13152.07231.94940.112.764.2117.05
AlexNet188226.29678.29972.904195.043.004.3018.54
ResNet18698766.541953.263137.8612049.882.554.0915.72
ResNet3412421309.223526.125759.1620966.752.694.4016.01
ResNet5017021914.965002.838369.4329597.102.614.3715.46
ResNet10133173672.619633.2915683.1657240.202.624.2715.59
ResNet15249325407.5813970.8823074.6882072.542.584.2715.18
ViT-H/1434204013.1811146.3117633.0766723.582.784.3916.63
Swin-B28813595.349505.3115054.8857310.032.644.1915.94
Average2.694.2816.23

Tree Flatten with Path

ModuleNodesOpTree (μs)JAX XLA (μs)PyTorch (μs)DM-Tree (μs)Speedup (J / O)Speedup (P / O)Speedup (D / O)
TinyMLP5336.49543.67N/A919.1314.90N/A25.19
AlexNet188115.442185.21N/A3752.1118.93N/A32.50
ResNet18698431.847106.55N/A12286.7016.46N/A28.45
ResNet341242845.6113431.99N/A22860.4815.88N/A27.03
ResNet5017021166.2718426.52N/A31225.0515.80N/A26.77
ResNet10133172312.7734770.49N/A59346.8615.03N/A25.66
ResNet15249323304.7450557.25N/A85847.9115.30N/A25.98
ViT-H/1434202235.2537473.53N/A64105.2416.76N/A28.68
Swin-B28811970.2532205.83N/A55177.5016.35N/A28.01
Average16.16N/A27.59

Tree Copy

ModuleNodesOpTree (μs)JAX XLA (μs)PyTorch (μs)DM-Tree (μs)Speedup (J / O)Speedup (P / O)Speedup (D / O)
TinyMLP5389.81232.26845.20981.482.599.4110.93
AlexNet188334.58959.323360.464316.052.8710.0412.90
ResNet186981128.112840.7111471.0712297.072.5210.1710.90
ResNet3412422160.575333.1020563.0621901.912.479.5210.14
ResNet5017022746.846823.8829705.9928927.882.4810.8110.53
ResNet10133175762.0513481.4556968.7860115.932.349.8910.43
ResNet15249328151.2120805.6181024.0684079.572.559.9410.31
ViT-H/1434205963.6115665.9159813.5268377.822.6310.0311.47
Swin-B28815401.5914255.3353361.7762317.072.649.8811.54
Average2.569.9711.02

Tree Map

ModuleNodesOpTree (μs)JAX XLA (μs)PyTorch (μs)DM-Tree (μs)Speedup (J / O)Speedup (P / O)Speedup (D / O)
TinyMLP5395.13243.86867.341026.992.569.1210.80
AlexNet188348.44987.573398.324354.812.839.7512.50
ResNet186981190.622982.6611719.9412559.012.519.8410.55
ResNet3412422205.875417.6020935.7222308.512.469.4910.11
ResNet5017023128.487579.5530372.7131638.672.429.7110.11
ResNet10133176173.0514846.5759167.8560245.422.419.589.76
ResNet15249328641.2222000.7484018.6586182.212.559.729.97
ViT-H/1434206211.7917077.4959790.2569763.862.759.6311.23
Swin-B28815673.6614339.6953309.1759764.612.539.4010.53
Average2.569.5810.62

Tree Map (nargs)

ModuleNodesOpTree (μs)JAX XLA (μs)PyTorch (μs)DM-Tree (μs)Speedup (J / O)Speedup (P / O)Speedup (D / O)
TinyMLP53137.06389.96N/A3908.772.85N/A28.52
AlexNet188467.241496.96N/A15395.133.20N/A32.95
ResNet186981603.794534.01N/A50323.762.83N/A31.38
ResNet3412422907.648435.33N/A90389.232.90N/A31.09
ResNet5017024183.7711382.51N/A121777.012.72N/A29.11
ResNet10133177721.1322247.85N/A238755.172.88N/A30.92
ResNet152493211508.0531429.39N/A360257.742.73N/A31.30
ViT-H/1434208294.2024524.86N/A270514.872.96N/A32.61
Swin-B28817074.6220854.80N/A241120.412.95N/A34.08
Average2.89N/A31.33

Tree Map with Path

ModuleNodesOpTree (μs)JAX XLA (μs)PyTorch (μs)DM-Tree (μs)Speedup (J / O)Speedup (P / O)Speedup (D / O)
TinyMLP53109.82778.30N/A2186.407.09N/A19.91
AlexNet188365.162939.36N/A8355.378.05N/A22.88
ResNet186981308.269529.58N/A25758.247.28N/A19.69
ResNet3412422527.2118084.89N/A45942.327.16N/A18.18
ResNet5017023226.0322935.53N/A61275.347.11N/A18.99
ResNet10133176663.5246878.89N/A126642.147.04N/A19.01
ResNet15249329378.1966136.44N/A176981.017.05N/A18.87
ViT-H/1434207033.6950418.37N/A142508.117.17N/A20.26
Swin-B28816078.1543173.22N/A116612.717.10N/A19.19
Average7.23N/A19.66

Tree Map with Path (nargs)

ModuleNodesOpTree (μs)JAX XLA (μs)PyTorch (μs)DM-Tree (μs)Speedup (J / O)Speedup (P / O)Speedup (D / O)
TinyMLP53146.05917.00N/A3940.616.28N/A26.98
AlexNet188489.273560.76N/A15434.717.28N/A31.55
ResNet186981712.7911171.44N/A50219.866.52N/A29.32
ResNet3412423112.8321024.58N/A95505.716.75N/A30.68
ResNet5017024220.7026600.82N/A121897.576.30N/A28.88
ResNet10133178631.3454372.37N/A236555.546.30N/A27.41
ResNet152493212710.4977643.13N/A353600.326.11N/A27.82
ViT-H/1434208753.0958712.71N/A286365.366.71N/A32.72
Swin-B28817359.2950112.23N/A228866.666.81N/A31.10
Average6.56N/A29.61

Changelog

See CHANGELOG.md.

License

OpTree is released under the Apache License 2.0.

OpTree is heavily based on JAX's implementation of the PyTree utility, with deep refactoring and several improvements. The original licenses can be found at JAX's Apache License 2.0 and Tensorflow's Apache License 2.0.