File size: 10,092 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
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
# Copyright 2019 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.
# ==============================================================================
"""Gradient clipping transformations.

Note that complex numbers are also supported, see
https://gist.github.com/wdphy16/118aef6fb5f82c49790d7678cf87da29
"""

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

from optax import tree_utils as otu
from optax._src import base
from optax._src import linear_algebra
from optax._src import numerics


def clip(max_delta: chex.Numeric) -> base.GradientTransformation:
  """Clips updates element-wise, to be in ``[-max_delta, +max_delta]``.

  Args:
    max_delta: The maximum absolute value for each element in the update.

  Returns:
    A `GradientTransformation` object.
  """

  def update_fn(updates, state, params=None):
    del params
    return otu.tree_clip(updates, -max_delta, max_delta), state

  return base.GradientTransformation(base.init_empty_state, update_fn)


def clip_by_block_rms(threshold: float) -> base.GradientTransformation:
  """Clips updates to a max rms for the gradient of each param vector or matrix.

  A `block` is here a weight vector (e.g. in a Linear layer) or a weight matrix
  (e.g. in a convolutional layer) appearing as a leaf in the grads/param pytree.

  Args:
    threshold: The maximum rms for the gradient of each param vector or matrix.

  Returns:
    A `GradientTransformation` object.
  """

  def update_fn(updates, state, params=None):
    del params

    def _clip_fn(u):
      clip_denom = jnp.maximum(
          1.0,
          jnp.sqrt(jnp.mean(numerics.abs_sq(u))) / threshold)
      return u / clip_denom

    updates = jtu.tree_map(_clip_fn, updates)
    return updates, state

  return base.GradientTransformation(base.init_empty_state, update_fn)


def clip_by_global_norm(max_norm: float) -> base.GradientTransformation:
  """Clips updates using their global norm.

  References:
    [Pascanu et al, 2012](https://arxiv.org/abs/1211.5063)

  Args:
    max_norm: The maximum global norm for an update.

  Returns:
    A `GradientTransformation` object.
  """

  def update_fn(updates, state, params=None):
    del params
    g_norm = linear_algebra.global_norm(updates)
    # TODO(b/163995078): revert back to the following (faster) implementation
    # once analysed how it affects backprop through update (e.g. meta-gradients)
    # g_norm = jnp.maximum(max_norm, g_norm)
    # updates = jtu.tree_map(lambda t: (t / g_norm) * max_norm, updates)
    trigger = jnp.squeeze(g_norm < max_norm)
    chex.assert_shape(trigger, ())  # A scalar.

    def clip_fn(t):
      return jax.lax.select(trigger, t, (t / g_norm.astype(t.dtype)) * max_norm)

    updates = jtu.tree_map(clip_fn, updates)
    return updates, state

  return base.GradientTransformation(base.init_empty_state, update_fn)


def per_example_global_norm_clip(
    grads: list[chex.Array], l2_norm_clip: float
) -> tuple[list[chex.Array], jax.Array]:
  """Applies gradient clipping per-example using their global norm.

  References:
    [Abadi et al, 2016](https://arxiv.org/abs/1607.00133)

  Args:
    grads: flattened update; the function expects these to have a batch
      dimension on the 0th axis.
    l2_norm_clip: maximum L2 norm of the per-example gradients.

  Returns:
    A tuple containing sum of the clipped per-example grads, and the number of
    per-example grads that were clipped.
  """
  bsize = grads[0].shape[0]

  if any(g.ndim == 0 or bsize != g.shape[0] for g in grads):
    raise ValueError(
        'Unlike other transforms, `per_example_global_norm_clip` expects'
        ' `grads` to have a batch dimension in the 0th axis.')

  global_grad_norms = jax.vmap(linear_algebra.global_norm)(grads)
  divisors = jnp.maximum(global_grad_norms / l2_norm_clip, 1.0)
  num_clipped = jnp.greater(divisors, 1.0).sum()
  clipped_sum = [(jnp.moveaxis(g, 0, -1) / divisors).sum(-1) for g in grads]
  return clipped_sum, num_clipped


def per_example_layer_norm_clip(
    grads: list[chex.Array],
    global_l2_norm_clip: float,
    uniform: bool = True,
    eps: float = 1e-8,
) -> tuple[list[chex.Array], list[chex.Array]]:
  """Applies gradient clipping per-example using per-layer norms.

  References:
    [McMahan et al, 2012](https://arxiv.org/abs/1710.06963)]

  Args:
    grads: flattened update; i.e. a list of gradients in which each item is
      the gradient for one layer; the function expects these to have a batch
      dimension on the 0th axis.
    global_l2_norm_clip: overall L2 clip norm to use.
    uniform: If `True`, per-layer clip norm is global_l2_norm_clip/sqrt(L),
      where L is the number of layers. Otherwise, per-layer clip norm is
      global_l2_norm_clip * sqrt(f), where f is the fraction of total model
      parameters that are in this layer.
    eps: Small positive value to add to norms to avoid possible division by
      zero.

  Let C = `global_l2_norm_clip value`. Then per-layer clipping is done as
  follows:
  (1) If `uniform` is `True`, each of the K layers has an individual clip
      norm of C / sqrt(K).
  (2) If `uniform` is `False`, each of the K layers has an individual clip
      norm of C * sqrt(D_i / D) where D_i is the number of parameters in
      layer i, and D is the total number of parameters in the model.

  Returns:
    A tuple containing sum of the clipped per-example grads and the number of
    per-example grads that were clipped for each layer.
  """
  bsize = grads[0].shape[0]

  if any(g.ndim == 0 or bsize != g.shape[0] for g in grads):
    raise ValueError(
        'Unlike other transforms, `per_example_layer_norm_clip` expects'
        ' `grads` to have a batch dimension in the 0th axis; got shapes:'
        f' {(g.shape for g in grads)}.'
    )

  num_layers = len(grads)

  # Compute per-layer clip norms, based on whether we are using uniform
  # variant or not.
  if uniform:
    # Create list of length `num_layers` of per-layer clip norm.
    layer_clip_norms = (
        global_l2_norm_clip * (1.0 / num_layers) ** 0.5,
    ) * num_layers
  else:
    total_params = sum(g[0].size for g in grads)
    layer_clip_norms = tuple(
        global_l2_norm_clip * (g[0].size / float(total_params)) ** 0.5
        for g in grads
    )

  # Compute per-layer grad norms.
  def map_layer_norm(grads_list):
    return [jnp.linalg.norm(g, ord=None, axis=None) for g in grads_list]

  layer_grad_norms_per_example = jax.vmap(map_layer_norm)(grads)

  # Perform clipping.
  divisors = (
      tuple(
          jnp.maximum(
              layer_grad_norm / (layer_clip_norm + eps), 1.0
          )
          for layer_grad_norm, layer_clip_norm in zip(
              layer_grad_norms_per_example, layer_clip_norms
          )
      )
  )
  num_clipped = [jnp.greater(divisor, 1.0).sum() for divisor in divisors]
  clipped_sum = [
      (g / jnp.expand_dims(d, axis=[i for i in range(1, g.ndim)])).sum(0)
      for g, d in zip(grads, divisors)
  ]
  return clipped_sum, num_clipped


def unitwise_norm(x: chex.Array) -> chex.Array:
  """Computes norms of each output unit separately."""
  if jnp.squeeze(x).ndim <= 1:  # Scalars and vectors
    squared_norm = jnp.sum(numerics.abs_sq(x), keepdims=True)
  # Note that this assumes parameters with a shape of length 3 are multihead
  # linear parameters--if you wish to apply AGC to 1D convs, you may need
  # to modify this line.
  elif x.ndim in (2, 3):  # Linear layers of shape IO or multihead linear
    squared_norm = jnp.sum(numerics.abs_sq(x), axis=0, keepdims=True)
  elif x.ndim == 4:  # Conv kernels of shape HWIO
    squared_norm = jnp.sum(numerics.abs_sq(x), axis=(0, 1, 2), keepdims=True)
  else:
    raise ValueError(
        f'Expected parameter with shape in {1, 2, 3, 4}, got {x.shape}.')
  chex.assert_is_broadcastable(squared_norm.shape, x.shape)
  return jnp.broadcast_to(jnp.sqrt(squared_norm), x.shape)


def unitwise_clip(g_norm: chex.Array,
                  max_norm: chex.Array,
                  grad: chex.Array,
                  div_eps: float = 1e-6) -> chex.Array:
  """Applies gradient clipping unit-wise."""
  # This little max(., div_eps) is distinct from the normal eps and just
  # prevents division by zero. It technically should be impossible to engage.
  clipped_grad = grad * (max_norm / jnp.maximum(g_norm, div_eps))
  chex.assert_equal_shape((g_norm, max_norm, grad, clipped_grad))
  return jnp.where(g_norm < max_norm, grad, clipped_grad)


def adaptive_grad_clip(clipping: float,
                       eps: float = 1e-3) -> base.GradientTransformation:
  """Clips updates to be at most ``clipping * parameter_norm``, unit-wise.

  References:
    [Brock, Smith, De, Simonyan 2021] High-Performance Large-Scale Image
    Recognition Without Normalization. (https://arxiv.org/abs/2102.06171)

  Args:
    clipping: The maximum allowed ratio of update norm to parameter norm.
    eps: An epsilon term to prevent clipping of zero-initialized params.

  Returns:
    A `GradientTransformation` object.
  """

  def update_fn(updates, state, params):
    if params is None:
      raise ValueError(base.NO_PARAMS_MSG)
    g_norm, p_norm = jtu.tree_map(unitwise_norm, (updates, params))
    # Maximum allowable norm.
    max_norm = jtu.tree_map(
        lambda x: clipping * jnp.maximum(x, eps), p_norm)
    # If grad norm > clipping * param_norm, rescale.
    updates = jtu.tree_map(unitwise_clip, g_norm, max_norm, updates)
    return updates, state

  return base.GradientTransformation(base.init_empty_state, update_fn)