feat(ops): add tree flatten and tree map functions with extra paths (#11

)

.github/workflows/tests.yml

@@ -85,7 +85,7 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-latest, windows-latest, macos-latest]
-        python-version: ["3.6", "3.8", "3.9", "3.10", "3.11"]
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
       fail-fast: false
     timeout-minutes: 30
     steps:

.pylintrc

@@ -53,7 +53,7 @@ persistent=yes
 
 # Minimum Python version to use for version dependent checks. Will default to
 # the version used to run pylint.
-py-version=3.6
+py-version=3.7
 
 # Discover python modules and packages in the file system subtree.
 recursive=no

README.md

@@ -1,6 +1,6 @@
 # OpTree
 
-![Python 3.6+](https://img.shields.io/badge/Python-3.6%2B-brightgreen)
+![Python 3.7+](https://img.shields.io/badge/Python-3.7%2B-brightgreen)
 [![PyPI](https://img.shields.io/pypi/v/optree?logo=pypi)](https://pypi.org/project/optree)
 ![GitHub Workflow Status](https://img.shields.io/github/workflow/status/metaopt/optree/Build?label=build&logo=github)
 ![GitHub Workflow Status](https://img.shields.io/github/workflow/status/metaopt/optree/Tests?label=tests&logo=github)
@@ -61,7 +61,7 @@ cd optree
 pip3 install .
 ```
 
-Compiling from the source requires Python 3.6+, a compiler (`gcc` / `clang` / `icc` / `cl.exe`) supports C++20 and a `cmake` installation.
+Compiling from the source requires Python 3.7+, a compiler (`gcc` / `clang` / `icc` / `cl.exe`) supports C++20 and a `cmake` installation.
 
 --------------------------------------------------------------------------------
 
@@ -122,17 +122,19 @@ The `NoneType` is a special case discussed in section [`None` is non-leaf Node v
 
 A container-like Python type can be registered in the container registry with a pair of functions that specify:
 
-- `flatten_func(container) -> (children, metadata)`: convert an instance of the container type to a `(children, metadata)` pair, where `children` is an iterable of subtrees.
+- `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)`.
+
 ```python
 >>> import torch
 
 >>> optree.register_pytree_node(
 ...     torch.Tensor,
-...     flatten_func=lambda tensor: (
+...     flatten_func=lambda tensor: (  # -> (children, metadata)
 ...         (tensor.cpu().numpy(),),
 ...         dict(dtype=tensor.dtype, device=tensor.device, requires_grad=tensor.requires_grad),
 ...     ),
@@ -154,6 +156,9 @@ The `metadata` is some necessary data apart from the children to reconstruct the
     })
 )
 
+>>> optree.tree_paths(tree)  # entries are not defined and use `range(len(children))`
+[('bias', 0), ('weight', 0)]
+
 >>> optree.tree_unflatten(treespec, leaves)
 {'bias': tensor([0., 0.]), 'weight': tensor([[1., 1.]], device='cuda:0')}
 ```
@@ -165,16 +170,24 @@ Users can also extend the pytree registry by decorating the custom class and def
 ...
 ... @optree.register_pytree_node_class
 ... class MyDict(UserDict):
-...     def tree_flatten(self):
+...     def tree_flatten(self):  # -> (children, metadata, entries)
 ...         reversed_keys = sorted(self.keys(), reverse=True)
-...         return [self[key] for key in reversed_keys], reversed_keys
+...         return (
+...             [self[key] for key in reversed_keys],  # children
+...             reversed_keys,  # metadata
+...             reversed_keys,  # entries
+...         )
 ...
 ...     @classmethod
 ...     def tree_unflatten(cls, metadata, children):
 ...         return MyDict(zip(metadata, children))
 
->>> optree.tree_flatten(MyDict(b=2, a=1, c=3))
-([3, 2, 1], PyTreeSpec(CustomTreeNode(MyDict[['c', 'b', 'a']], [*, *, *])))
+>>> optree.tree_flatten_with_path(MyDict(b=4, a=(2, 3), c=MyDict({'d': 5, 'f': 6})))
+(
+    [('c', 'f'), ('c', 'd'), ('b',), ('a', 0), ('a', 1)],
+    [6, 5, 4, 2, 3],
+    PyTreeSpec(CustomTreeNode(MyDict[['c', 'b', 'a']], [CustomTreeNode(MyDict[['f', 'd']], [*, *]), *, (*, *)]))
+)
 ```
 
 #### Limitations of the PyTree Type Registry
@@ -195,21 +208,21 @@ It may also serve as a sentinel value.
 In addition, if a function has returned without any return value, it also implicitly returns 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 slightly different than the definition of a non-leaf node as discussed above.
+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.
 
 ```python
 >>> 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(NoneIsLeaf, {'a': *, 'b': (*, [*, *]), 'c': *, 'd': *}))
+([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, *))
+([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.

include/registry.h

@@ -52,7 +52,7 @@ class PyTreeTypeRegistry {
         // The following values are populated for custom types.
         // The Python type object, used to identify the type.
         py::object type;
-        // A function with signature: object -> (iterable, metadata)
+        // A function with signature: object -> (iterable, metadata, entries)
         py::function to_iterable;
         // A function with signature: (metadata, iterable) -> object
         py::function from_iterable;

include/treespec.h

@@ -28,6 +28,7 @@ limitations under the License.
 #include <optional>
 #include <stdexcept>
 #include <string>
+#include <tuple>
 #include <utility>
 #include <vector>
 
@@ -64,6 +65,26 @@ class PyTreeSpec {
                      const std::optional<py::function> &leaf_predicate = std::nullopt,
                      const bool &none_is_leaf = false);
 
+    // Flattens a PyTree into a list of leaves with a list of paths and a PyTreeSpec.
+    // Returns references to the flattened objects, which might be temporary objects in the case of
+    // custom PyType handlers.
+    static std::tuple<std::vector<py::object>, std::vector<py::object>, std::unique_ptr<PyTreeSpec>>
+    FlattenWithPath(const py::handle &tree,
+                    const std::optional<py::function> &leaf_predicate = std::nullopt,
+                    const bool &none_is_leaf = false);
+
+    // Recursive helper used to implement FlattenWithPath().
+    void FlattenIntoWithPath(const py::handle &handle,
+                             std::vector<py::object> &leaves,  // NOLINT
+                             std::vector<py::object> &paths,   // NOLINT
+                             const std::optional<py::function> &leaf_predicate = std::nullopt,
+                             const bool &none_is_leaf = false);
+    void FlattenIntoWithPath(const py::handle &handle,
+                             absl::InlinedVector<py::object, 2> &leaves,  // NOLINT
+                             absl::InlinedVector<py::object, 2> &paths,   // NOLINT
+                             const std::optional<py::function> &leaf_predicate = std::nullopt,
+                             const bool &none_is_leaf = false);
+
     // Flattens a PyTree up to this PyTreeSpec. 'this' must be a tree prefix of the tree-structure
     // of 'x'. For example, if we flatten a value [(1, (2, 3)), {"foo": 4}] with a PyTreeSpec [(*,
     // *), *], the result is the list of leaves [1, (2, 3), {"foo": 4}].
@@ -146,6 +167,13 @@ class PyTreeSpec {
         // For a Custom type, contains the auxiliary data returned by the `to_iterable` function.
         py::object node_data;
 
+        // The tuple of path entries.
+        // This is optional, if not specified, `range(arity)` is used.
+        // For a sequence, contains the index of the element.
+        // For a mapping, contains the key of the element.
+        // For a Custom type, contains the path entries returned by the `to_iterable` function.
+        py::object node_entries;
+
         // Custom type registration. Must be null for non-custom types.
         const PyTreeTypeRegistry::Registration *custom = nullptr;
 
@@ -176,6 +204,14 @@ class PyTreeSpec {
                          Span &leaves,  // NOLINT
                          const std::optional<py::function> &leaf_predicate);
 
+    template <bool NoneIsLeaf, typename Span, typename Stack>
+    void FlattenIntoWithPathImpl(const py::handle &handle,
+                                 Span &leaves,  // NOLINT
+                                 Span &paths,   // NOLINT
+                                 Stack &stack,  // NOLINT
+                                 const ssize_t &depth,
+                                 const std::optional<py::function> &leaf_predicate);
+
     py::list FlattenUpToImpl(const py::handle &full_tree) const;
 
     template <bool NoneIsLeaf>

include/utils.h

@@ -28,6 +28,7 @@ limitations under the License.
 #include <optional>
 #include <stdexcept>
 #include <string>
+#include <tuple>
 #include <utility>
 #include <vector>
 

optree/_C.pyi

@@ -16,10 +16,10 @@
 # pylint: disable=all
 # isort: off
 
-from typing import TYPE_CHECKING, Callable, Iterable, List, Optional, Sequence, Tuple, Type
+from typing import TYPE_CHECKING, Any, Callable, Iterable, List, Optional, Sequence, Tuple, Type
 
 if TYPE_CHECKING:
-    from optree.typing import MetaData, Children, CustomTreeNode, PyTree, T, U
+    from optree.typing import Children, CustomTreeNode, MetaData, PyTree, T, U
 
 version: int
 
@@ -28,6 +28,11 @@ def flatten(
     leaf_predicate: Optional[Callable[[T], bool]] = None,
     node_is_leaf: bool = False,
 ) -> Tuple[List[T], 'PyTreeSpec']: ...
+def flatten_with_path(
+    tree: PyTree[T],
+    leaf_predicate: Optional[Callable[[T], bool]] = None,
+    node_is_leaf: bool = False,
+) -> Tuple[List[Tuple[Any, ...]], List[T], 'PyTreeSpec']: ...
 def all_leaves(iterable: Iterable[T], node_is_leaf: bool = False) -> bool: ...
 def leaf(node_is_leaf: bool = False) -> 'PyTreeSpec': ...
 def none(node_is_leaf: bool = False) -> 'PyTreeSpec': ...
@@ -53,7 +58,7 @@ class PyTreeSpec:
     def __hash__(self) -> int: ...
 
 def register_node(
-    type: Type[CustomTreeNode[T]],
+    cls: Type[CustomTreeNode[T]],
     to_iterable: Callable[[CustomTreeNode[T]], Tuple[Children[T], MetaData]],
     from_iterable: Callable[[MetaData, Children[T]], CustomTreeNode[T]],
 ) -> None: ...

optree/__init__.py

@@ -22,9 +22,13 @@
     tree_all,
     tree_any,
     tree_flatten,
+    tree_flatten_with_path,
     tree_leaves,
     tree_map,
     tree_map_,
+    tree_map_with_path,
+    tree_map_with_path_,
+    tree_paths,
     tree_reduce,
     tree_replace_nones,
     tree_structure,
@@ -60,12 +64,16 @@
     'NONE_IS_NODE',
     'NONE_IS_LEAF',
     'tree_flatten',
+    'tree_flatten_with_path',
     'tree_unflatten',
     'tree_leaves',
     'tree_structure',
+    'tree_paths',
     'all_leaves',
     'tree_map',
     'tree_map_',
+    'tree_map_with_path',
+    'tree_map_with_path_',
     'tree_reduce',
     'tree_transpose',
     'tree_replace_nones',