Comparison Between TreeValue and DM-Tree

In this section, we will compare the feature and performance of the dm-tree library, which is developed by deepmind.

Before starting the comparison, let us define some thing.

origin = {'a': 1, 'b': 2, 'c': {'x': 3, 'y': 4}}

Mapping Operation

Mapping operation is quite common in the processing of trees. A mapping function should be provided to create a new tree based on the mapped tree.

TreeValue’s mapping

In TreeValue, mapping is provided to simply create a new tree.

from treevalue import FastTreeValue, mapping

tv = FastTreeValue(origin)
tv

mapping(tv, lambda x: x * 2)

Here is the performance test.

%timeit mapping(tv, lambda x: x * 2)

2.65 µs ± 51.1 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

In order to support the cased that the mapped value is related to both path and value of each node, we can use the ‘path mapping mode’ by simply use the second parameter of the mapping function.

mapping(tv, lambda x, p: ('path', p, 'value', x))

And here is the performance

%timeit mapping(tv, lambda x, p: ('path', p, 'value', x))

2.68 µs ± 19.6 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

DM-Tree’s mapping

In DM-Tree, mapping operation is supported by map_structure function.

from tree import map_structure

map_structure(lambda x: x * 2, origin)

{'a': 2, 'b': 4, 'c': {'x': 6, 'y': 8}}

This is the performance of map_structure, obviously much slower than mapping in TreeValue.

%timeit map_structure(lambda x: x * 2, origin)

18.5 µs ± 236 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

To supported the second situation in the last section, map_structure_with_path can be used.

from tree import map_structure_with_path

map_structure_with_path(lambda path, x: ('path', path, 'value', x), origin)

{'a': ('path', ('a',), 'value', 1),
 'b': ('path', ('b',), 'value', 2),
 'c': {'x': ('path', ('c', 'x'), 'value', 3),
  'y': ('path', ('c', 'y'), 'value', 4)}}

Here is the performance.

%timeit map_structure_with_path(lambda path, x: ('path', path, 'value', x), origin)

40.9 µs ± 944 ns per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

Flatten and Unflatten

In tree operations, flatten is often used to linearly expand the tree structure for operations such as parallel processing. Based on flatten, unflatten is its inverse operation, which can recover the tree structure from the linear data.

TreeValue’s Performance

In TreeValue, flatten and unflatten are provided, which usage are simple.

from treevalue import FastTreeValue, flatten, unflatten

origin_tree = FastTreeValue(origin)
origin_tree

flatted = flatten(origin_tree)
flatted

[(('a',), 1), (('b',), 2), (('c', 'x'), 3), (('c', 'y'), 4)]

Here is the performance of flatten

%timeit flatten(origin_tree)

838 ns ± 10.2 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

The tree can be re-created from flatted with function unflatten.

unflatten(flatted)

And here is the performance.

%timeit unflatten(flatted)

985 ns ± 9.63 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

DM-Tree’s Performance

Flatten is provided in DM-Tree as well, but it differs from that in TreeValue, as the following

from tree import flatten

flatten(origin)

[1, 2, 3, 4]

Here is the performance

%timeit flatten(origin)

1.22 µs ± 16.3 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

The structure of the tree is dropped, only the data is extracted from the tree. This means the flatten function in DM-Tree is irreversible, we can not recover the original tree with the result above.

The reversible resolution provided in DM-Tree is flatten_with_path

from tree import flatten_with_path

flatten_with_path(origin)

[(('a',), 1), (('b',), 2), (('c', 'x'), 3), (('c', 'y'), 4)]

Here is the performance, much slower than flatten in TreeValue.

%timeit flatten_with_path(origin)

19.3 µs ± 166 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

To re-create the original tree, we need a tree structure with any objects filled inside. Use the unflatten_as to archive this goal.

from tree import unflatten_as

unflatten_as({'a': None, 'b': None, 'c': {'x': None, 'y': None}}, [1, 2, 3, 4])

{'a': 1, 'b': 2, 'c': {'x': 3, 'y': 4}}

Here is the performance.

%timeit unflatten_as({'a': None, 'b': None, 'c': {'x': None, 'y': None}}, [1, 2, 3, 4])

15.4 µs ± 141 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

Positional Replacement

It is obvious that the unflatten_as in DM-Tree is quite different from unflatten in TreeValue, for the former one is replacing and the latter one is constructing. This means the unflatten_as in DM-Tree may supported some more features, such as creating a new tree with another tree’s structure and the given values. So we need an experiment on this, called ‘positional replacement’.

First, in TreeValue, we can build a function named replace to realize this.

from treevalue import flatten, unflatten

def replace(t, v):
    pairs = flatten(t)
    return unflatten([(path, vi) for (path, _), vi in zip(pairs, v)])

Create a new tree based on origin_tree’s structure and new values.

replace(origin_tree, [3, 5, 7, 9])

Here is the performance.

%timeit replace(origin_tree, [3, 5, 7, 9])

2.93 µs ± 43.9 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

In DM-Tree, unflatten_as can be directly used.

from tree import unflatten_as

unflatten_as(origin, [3, 5, 7, 9])

{'a': 3, 'b': 5, 'c': {'x': 7, 'y': 9}}

Here is the performance, even much slower than the replace function for integration.

%timeit unflatten_as(origin, [3, 5, 7, 9])

15 µs ± 95.4 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

Conclusion

The mapping operation is supported by both library, and mapping in TreeValue’s performance is significantly higher than the map_structure and map_structure_with_path in DM-Tree.

The flatten and unflatten in TreeValue is reversible, but in DM-Tree not. DM-Tree’s performance on flatten and unflatten operation is lower than that in TreeValue in all aspects.