File size: 4,717 Bytes
09d8e80
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
# Copyright 2021 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.
# ==============================================================================

"""Euclidean projections."""

from typing import Any

import chex

import jax
from jax import flatten_util
from jax import tree_util as jtu
import jax.numpy as jnp


def projection_non_negative(pytree: Any) -> Any:
  r"""Projection onto the non-negative orthant.

  .. math::

    \underset{p}{\text{argmin}} ~ ||x - p||_2^2 \quad
    \textrm{subject to} \quad p \ge 0

  where :math:`x` is the input pytree.

  Args:
    pytree: pytree to project.
  Returns:
    projected pytree, with the same structure as ``pytree``.
  """
  return jtu.tree_map(jax.nn.relu, pytree)


def _clip_safe(leaf, lower, upper):
  return jnp.clip(jnp.asarray(leaf), lower, upper)


def projection_box(pytree: Any, lower: Any, upper: Any) -> Any:
  r"""Projection onto box constraints.

  .. math::

    \underset{p}{\text{argmin}} ~ ||x - p||_2^2 \quad \textrm{subject to} \quad
    \text{lower} \le p \le \text{upper}

  where :math:`x` is the input pytree.

  Args:
    pytree: pytree to project.
    lower:  lower bound, a scalar or pytree with the same structure as
      ``pytree``.
    upper:  upper bound, a scalar or pytree with the same structure as
      ``pytree``.
  Returns:
    projected pytree, with the same structure as ``pytree``.
  """
  return jtu.tree_map(_clip_safe, pytree, lower, upper)


def projection_hypercube(pytree: Any, scale: Any = 1.0) -> Any:
  r"""Projection onto the (unit) hypercube.

  .. math::

    \underset{p}{\text{argmin}} ~ ||x - p||_2^2 \quad \textrm{subject to} \quad
    0 \le p \le \text{scale}

  where :math:`x` is the input pytree.

  By default, we project to the unit hypercube (`scale=1.0`).

  This is a convenience wrapper around
  :func:`projection_box <optax.projections.projection_box>`.

  Args:
    pytree: pytree to project.
    scale: scale of the hypercube, a scalar or a pytree (default: 1.0).
  Returns:
    projected pytree, with the same structure as ``pytree``.
  """
  return projection_box(pytree, lower=0.0, upper=scale)


@jax.custom_jvp
def _projection_unit_simplex(values: chex.Array) -> chex.Array:
  """Projection onto the unit simplex."""
  s = 1.0
  n_features = values.shape[0]
  u = jnp.sort(values)[::-1]
  cumsum_u = jnp.cumsum(u)
  ind = jnp.arange(n_features) + 1
  cond = s / ind + (u - cumsum_u / ind) > 0
  idx = jnp.count_nonzero(cond)
  return jax.nn.relu(s / idx + (values - cumsum_u[idx - 1] / idx))


@_projection_unit_simplex.defjvp
def _projection_unit_simplex_jvp(
    primals: list[chex.Array], tangents: list[chex.Array]
) -> tuple[chex.Array, chex.Array]:
  values, = primals
  values_dot, = tangents
  primal_out = _projection_unit_simplex(values)
  supp = primal_out > 0
  card = jnp.count_nonzero(supp)
  tangent_out = supp * values_dot - (jnp.dot(supp, values_dot) / card) * supp
  return primal_out, tangent_out


def projection_simplex(pytree: Any,
                       scale: chex.Numeric = 1.0) -> Any:
  r"""Projection onto a simplex.

  This function solves the following constrained optimization problem,
  where ``p`` is the input pytree.

  .. math::

    \underset{p}{\text{argmin}} ~ ||x - p||_2^2 \quad \textrm{subject to} \quad
    p \ge 0, p^\top 1 = \text{scale}

  By default, the projection is onto the probability simplex (unit simplex).

  Args:
    pytree: pytree to project.
    scale: value the projected pytree should sum to (default: 1.0).
  Returns:
    projected pytree, a pytree with the same structure as ``pytree``.

  .. versionadded:: 0.2.3

  Example:

    Here is an example using a pytree::

      >>> import jax.numpy as jnp
      >>> from optax import tree_utils, projections
      >>> pytree = {"w": jnp.array([2.5, 3.2]), "b": 0.5}
      >>> tree_utils.tree_sum(pytree)
      6.2
      >>> new_pytree = projections.projection_simplex(pytree)
      >>> tree_utils.tree_sum(new_pytree)
      1.0000002
  """
  if scale is None:
    scale = 1.0

  values, unravel_fn = flatten_util.ravel_pytree(pytree)
  new_values = scale * _projection_unit_simplex(values / scale)

  return unravel_fn(new_values)