File size: 8,346 Bytes
74e8f2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2024 Big Vision Authors.
#
# 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 transformations and other optax utilities."""

import operator
import big_vision.utils as u
import jax
import jax.numpy as jnp
import optax


def find_states(opt_state, cls):
  leaves = jax.tree.leaves(
      opt_state, is_leaf=lambda node: isinstance(node, cls))
  return [leaf for leaf in leaves if isinstance(leaf, cls)]


def get_count(opt_state, jittable=False):
  """Returns `ScaleByScheduleState.count` from `opt_state` as an integer."""
  counts = [
      state.count
      for state in find_states(opt_state, optax.ScaleByScheduleState)
  ]
  if jittable:
    return counts[0]
  else:
    counts = {int(c) for c in counts}
    assert len(counts) == 1, f"Expected exactly 1 ScaleByScheduleState:{counts}"
    return next(iter(counts))


def replace_frozen(schedule, pytree, replacement, log=None):
  """Replaces values matching frozen params in `pytree` with `replacement`."""
  if not isinstance(schedule, (list, tuple)):
    return pytree
  masks, scheds = _make_mask_trees(pytree, schedule, log=log)
  frozen_mask, _, _ = _split_frozen(masks, scheds)
  return jax.tree.map(
      lambda v, f: replacement if f else v, pytree, frozen_mask)


def clip_by_per_example_global_norm(
    max_norm: float,
) -> optax.GradientTransformation:
  """Clips the norm of per-example gradients."""

  def init_fn(params):
    del params
    return optax.EmptyState()

  def update_fn(updates, state, params=None):
    del params
    grads_flat, grads_treedef = jax.tree_util.tree_flatten(updates)
    clipped, _ = optax.per_example_global_norm_clip(grads_flat, max_norm)
    return jax.tree_util.tree_unflatten(grads_treedef, clipped), state

  return optax.GradientTransformation(init_fn, update_fn)


def make(config, params, *, sched_kw):
  """Returns gradient transform and learning rate functions."""

  # Global schedule. No schedule means frozen.
  schedule = config.get("schedule", {})
  if not isinstance(schedule, (tuple, list)):
    schedule = [(".*", schedule)]
  masks, scheds = _make_mask_trees(params, schedule, "config.schedule")
  frozen_mask, masks, scheds = _split_frozen(masks, scheds)
  not_frozen_mask = jax.tree.map(operator.not_, frozen_mask)
  def create_schedule(mult=1.0, **kw):
    assert "base" not in kw, kw
    return u.create_learning_rate_schedule(base=mult, **kw)
  schedule_fns = [create_schedule(**sched_kw, **sched) for sched in scheds]
  schedule_txs = [
      optax.masked(optax.scale_by_schedule(schedule_fn), mask)
      for schedule_fn, mask in zip(schedule_fns, masks)
  ] + [
      # Removes weight decay updates. Note that weight decay already has an
      # independent mask (which cannot be combined easily with a second mask),
      # so instead we multiply updates for frozen params with zero.
      optax.masked(optax.set_to_zero(), frozen_mask)
  ]

  # Gradient clipping.
  if clip_norm := config.get("grad_clip_norm"):
    if config.get("grad_clip_per_example"):
      clip_tx = clip_by_per_example_global_norm(clip_norm)
    else:
      clip_tx = optax.clip_by_global_norm(clip_norm)
    grad_clip_norm_tx = optax.masked(clip_tx, not_frozen_mask)
  else:
    grad_clip_norm_tx = optax.identity()

  # Optimizer updates.
  tx_func = operator.attrgetter(config.optax_name)(optax)
  opt_txs = [optax.masked(tx_func(**config.get("optax", {})), not_frozen_mask)]
  assert "optim" not in config, "Deprecated option, use config.optax."

  # Learning rate multipliers. Defaults to 1.0.
  lr_mult_txs = [optax.scale(config.lr)]
  if config.get("lr_mults"):
    masks, mults = _make_mask_trees(params, config.lr_mults, "config.lr_mults")
    assert all(mult > 0 for mult in mults), (
        f"Use schedule=None for parameter freezing instead of lr_mults={mults}")
    lr_mult_txs += [
        optax.masked(optax.scale(mult), mask)
        for mult, mask in zip(mults, masks)
    ]

  # Weight decay. Defaults to 0.0.
  # Weight decay is not gradient-based but instead uses "params side-input".
  # Hence, weight decay is additive and independent of previous gradient-based
  # updates.
  assert "weight_decay" not in config, "Deprecated option. Use wd and schedule."
  assert config.get("weight_decay_decouple", True), (
      "Coupled weight decay not supported anymore.")
  if config.get("wd"):
    wd_mults = config.get("wd_mults", [(".*/kernel$", 1.0)])
    masks, mults = _make_mask_trees(params, wd_mults, "config.wd_mults")
    weight_decay_txs = [
        optax.add_decayed_weights(config.wd * mult, mask)
        for mult, mask in zip(mults, masks)
    ]
  else:
    weight_decay_txs = []

  # Combine gradient updates and learning rate schedules.
  return optax.chain(
      grad_clip_norm_tx,
      *opt_txs,
      *lr_mult_txs,
      *weight_decay_txs,
      *schedule_txs,
      optax.scale(-1.0)), schedule_fns


def _make_mask_trees(params, patterns_values, log):
  patterns, values = zip(*patterns_values)
  masks = u.make_mask_trees(params, patterns, log=log)
  return masks, values


def _split_frozen(masks, scheds):
  """Computes `frozen_mask` and updates `masks` and `scheds`."""
  # Specifying `None` as a scheduler freezes params.
  all_false = jax.tree.map(lambda *bools: not any(bools), *masks)
  not_covered = [k for k, v in u.tree_flatten_with_names(all_false)[0] if v]
  assert not not_covered, (
      f"All params must be covered (use `None` for freezing): {not_covered}")
  frozen_masks = [
      mask for mask, sched in zip(masks, scheds) if sched is None]
  frozen_mask = jax.tree.map(
      lambda *bools: any(bools), *frozen_masks,
      all_false)  # `all_false` is required when `frozen_masks==[]`.
  masks, scheds = zip(*(
      (mask, sched) for mask, sched in zip(masks, scheds) if sched is not None))
  return frozen_mask, masks, scheds


############ Custom BigVision optimizers #######################################
# Currently there's only one custom optimizer and we don't foresee new ones in
# the near future, we opt not to create a new optimizer folder/module for just
# one isolated case. If there will be more optimizers, we can consider moving
# them into individual files in a subfolder.


# A dummy object to allow for foo.bar access syntax, see
# https://stackoverflow.com/a/19476841/2366315
optax.big_vision = type("", (), {})()


def scale_by_adafactor(min_dim_size_to_factor=32,
                       decay_rate=0.8, decay_offset=0,
                       beta2_cap=0.999,
                       clipping_threshold=None,
                       momentum=0.9, dtype_momentum=jnp.bfloat16,
                       eps=1e-30):
  """The BigVision variant of Adafactor optimizer."""

  def _decay_rate_pow(i, exponent):
    """Second-order moment decay schedule."""
    t = jnp.array(i, jnp.float32) + 1.0
    return jnp.minimum(beta2_cap, 1.0 - t**(-exponent))

  scale_by_rms = optax.scale_by_factored_rms(
      factored=True,
      decay_rate=decay_rate,
      step_offset=decay_offset,
      min_dim_size_to_factor=min_dim_size_to_factor,
      epsilon=eps,
      decay_rate_fn=_decay_rate_pow)

  clip = (optax.clip_by_block_rms(clipping_threshold) if clipping_threshold
          else optax.identity())

  mom = (optax.ema(momentum, debias=False, accumulator_dtype=dtype_momentum)
         if momentum else optax.identity())

  return optax.chain(scale_by_rms, clip, mom)

optax.big_vision.scale_by_adafactor = scale_by_adafactor  # pytype: disable=module-attr


# A few more aliases we use frequently:
def momentum_hp(momentum=0.9, dtype=jnp.bfloat16, nesterov=False):
  """SGD-Momentum with half-precision accumulator."""
  return optax.trace(decay=momentum, accumulator_dtype=dtype, nesterov=nesterov)

optax.big_vision.momentum_hp = momentum_hp  # pytype: disable=module-attr
optax.big_vision.sgd = optax.identity  # pytype: disable=module-attr