File size: 19,629 Bytes
eb49b41
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
# coding=utf-8
# Copyright (c) 2019 NVIDIA CORPORATION. All rights reserved.
# Copyright 2018 The Google AI Language Team 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.

"""Functions and classes related to optimization (weight updates)."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import re
import tensorflow as tf
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import linalg_ops
from tensorflow.python.ops import math_ops

try:
    import horovod.tensorflow as hvd
except ImportError:
    hvd = None

def horovod_enabled():
  return hvd is not None and hvd.is_initialized()

def create_optimizer(loss, init_lr, num_train_steps, num_warmup_steps, manual_fp16=False, use_fp16=False, num_accumulation_steps=1,
                     optimizer_type="adam", allreduce_post_accumulation=False, init_loss_scale=2**32, use_tpu=False):
  """Creates an optimizer training op."""
  global_step = tf.compat.v1.train.get_or_create_global_step()

  # avoid step change in learning rate at end of warmup phase
  if optimizer_type == "adam":
    power = 1.0
    decayed_learning_rate_at_crossover_point = init_lr * (
        (1.0 - float(num_warmup_steps) / float(num_train_steps)) ** power)
  else:
    power = 0.5
    decayed_learning_rate_at_crossover_point = init_lr

  adjusted_init_lr = init_lr * (init_lr / decayed_learning_rate_at_crossover_point)
  print('decayed_learning_rate_at_crossover_point = %e, adjusted_init_lr = %e' %
        (decayed_learning_rate_at_crossover_point, adjusted_init_lr))

  learning_rate = tf.constant(value=adjusted_init_lr, shape=[], dtype=tf.float32)

  # Implements linear decay of the learning rate.
  learning_rate = tf.compat.v1.train.polynomial_decay(
      learning_rate,
      global_step,
      num_train_steps,
      end_learning_rate=0.0,
      power=power,
      cycle=False)

  # Implements linear warmup. I.e., if global_step < num_warmup_steps, the
  # learning rate will be `global_step/num_warmup_steps * init_lr`.
  if num_warmup_steps:
    global_steps_int = tf.cast(global_step, tf.int32)
    warmup_steps_int = tf.constant(num_warmup_steps, dtype=tf.int32)

    global_steps_float = tf.cast(global_steps_int, tf.float32)
    warmup_steps_float = tf.cast(warmup_steps_int, tf.float32)

    warmup_percent_done = global_steps_float / warmup_steps_float
    warmup_learning_rate = init_lr * warmup_percent_done

    is_warmup = tf.cast(global_steps_int < warmup_steps_int, tf.float32)
    learning_rate = (
        (1.0 - is_warmup) * learning_rate + is_warmup * warmup_learning_rate)

  if optimizer_type == "lamb":
    print("Initializing LAMB Optimizer")
    optimizer = LAMBOptimizer(
        learning_rate=learning_rate,
        weight_decay_rate=0.01,
        beta_1=0.9,
        beta_2=0.999,
        epsilon=1e-6,
        exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"])
  else:
    print("Initializing ADAM Weight Decay Optimizer")
    # It is recommended that you use this optimizer for fine tuning, since this
    # is how the model was trained (note that the Adam m/v variables are NOT
    # loaded from init_checkpoint.)
    optimizer = AdamWeightDecayOptimizer(
        learning_rate=learning_rate,
        weight_decay_rate=0.01,
        beta_1=0.9,
        beta_2=0.999,
        epsilon=1e-6,
        exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"])

  if horovod_enabled() and (num_accumulation_steps == 1 or (not allreduce_post_accumulation)):
    optimizer = hvd.DistributedOptimizer(optimizer, sparse_as_dense=True)
  if use_fp16:
    loss_scaler = tf.train.experimental.DynamicLossScale(
        initial_loss_scale=init_loss_scale, increment_period=1000, multiplier=2.0)
    optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(optimizer, loss_scaler)
    loss_scale_value = tf.identity(loss_scaler(), name="loss_scale")
  if manual_fp16:
    assert False, "No support for ExponentialUpdateLossScaleManager and LossScaleOptimizer in TF2.0"
    loss_scale_manager = tf.contrib.mixed_precision.ExponentialUpdateLossScaleManager(init_loss_scale=init_loss_scale,
                                                                                      incr_every_n_steps=1000,
                                                                                      decr_every_n_nan_or_inf=2,
                                                                                      decr_ratio=0.5)
    optimizer = tf.contrib.mixed_precision.LossScaleOptimizer(optimizer, loss_scale_manager)
  if use_tpu:
    optimizer = tf.compat.v1.tpu.CrossShardOptimizer(optimizer)
  tvars = tf.compat.v1.trainable_variables()

  if num_accumulation_steps > 1:
    grads_and_vars = optimizer.compute_gradients(loss * 1.0 / num_accumulation_steps, tvars)
    local_step = tf.compat.v1.get_variable(name="local_step", shape=[], dtype=tf.int32, trainable=False,
                                           initializer=tf.compat.v1.zeros_initializer)
    batch_finite = tf.compat.v1.get_variable(name="batch_finite", shape=[], dtype=tf.bool, trainable=False,
                                             initializer=tf.compat.v1.ones_initializer)
    accum_vars = [tf.compat.v1.get_variable(
        name=tvar.name.split(":")[0] + "/accum",
        shape=tvar.shape.as_list(),
        dtype=tf.float32,
        trainable=False,
        initializer=tf.compat.v1.zeros_initializer()) for tvar in tf.compat.v1.trainable_variables()]

    reset_step = tf.cast(tf.math.equal(local_step % num_accumulation_steps, 0), dtype=tf.bool)
    local_step = tf.cond(pred=reset_step, true_fn=lambda: local_step.assign(
        tf.ones_like(local_step)), false_fn=lambda: local_step.assign_add(1))

    grads_and_vars_and_accums = [(gv[0], gv[1], accum_vars[i])
                                 for i, gv in enumerate(grads_and_vars) if gv[0] is not None]
    grads, tvars, accum_vars = list(zip(*grads_and_vars_and_accums))

    all_are_finite = tf.reduce_all(input_tensor=[tf.reduce_all(input_tensor=tf.math.is_finite(
        g)) for g in grads]) if manual_fp16 or use_fp16 else tf.constant(True, dtype=tf.bool)
    batch_finite = tf.cond(pred=reset_step,
                           true_fn=lambda: batch_finite.assign(tf.math.logical_and(
                               tf.constant(True, dtype=tf.bool), all_are_finite)),
                           false_fn=lambda: batch_finite.assign(tf.math.logical_and(batch_finite, all_are_finite)))

    # This is how the model was pre-trained.
    # ensure global norm is a finite number
    # to prevent clip_by_global_norm from having a hizzy fit.
    (clipped_grads, _) = tf.clip_by_global_norm(
        grads, clip_norm=1.0,
        use_norm=tf.cond(
            pred=all_are_finite,
            true_fn=lambda: tf.linalg.global_norm(grads),
            false_fn=lambda: tf.constant(1.0)))

    accum_vars = tf.cond(pred=reset_step,
                         true_fn=lambda: [accum_vars[i].assign(grad) for i, grad in enumerate(clipped_grads)],
                         false_fn=lambda: [accum_vars[i].assign_add(grad) for i, grad in enumerate(clipped_grads)])

    update_step = tf.identity(tf.cast(tf.math.equal(local_step % num_accumulation_steps, 0),
                                      dtype=tf.bool), name="update_step")

    def allreduce_of_batch_finite_required():
      # In case of bf16 and fp32 batch finite is tf.constant(True, dtype=tf.bool)
      return horovod_enabled() and manual_fp16 and use_fp16

    # TODO: in future if we want to enable infinite batch iter skiping we will need to change this allreduce.
    new_global_step = tf.cond(pred=tf.math.logical_and(update_step,
                                                       tf.cast(hvd.allreduce(tf.cast(batch_finite, tf.int32)), tf.bool) if allreduce_of_batch_finite_required() else batch_finite),
                              true_fn=lambda: global_step + 1,
                              false_fn=lambda: global_step)
    new_global_step = tf.identity(new_global_step, name='step_update')

    def update(accum_vars):
      with tf.control_dependencies([global_step.assign(new_global_step)]):
        if allreduce_post_accumulation and horovod_enabled():
          accum_vars = [hvd.allreduce(tf.convert_to_tensor(value=accum_var)) if isinstance(accum_var, tf.IndexedSlices)
                        else hvd.allreduce(accum_var) for accum_var in accum_vars]

        return optimizer.apply_gradients(list(zip(accum_vars, tvars)), global_step=global_step)

    train_op = tf.cond(pred=update_step,
                       true_fn=lambda: update(accum_vars), false_fn=lambda: tf.no_op())
  else:
    grads_and_vars = optimizer.compute_gradients(loss, tvars)
    if horovod_enabled():
      grads_and_vars = [(g, v) for g, v in grads_and_vars if g is not None]
      grads, tvars = list(zip(*grads_and_vars))
    else:
      grads = tf.gradients(ys=loss, xs=tvars)
    all_are_finite = tf.reduce_all(
        input_tensor=[tf.reduce_all(input_tensor=tf.math.is_finite(g)) for g in grads]) if use_fp16 or manual_fp16 else tf.constant(True, dtype=tf.bool)

    # This is how the model was pre-trained.
    # ensure global norm is a finite number
    # to prevent clip_by_global_norm from having a hizzy fit.
    (clipped_grads, _) = tf.clip_by_global_norm(
        grads, clip_norm=1.0,
        use_norm=tf.cond(
            pred=all_are_finite,
            true_fn=lambda: tf.linalg.global_norm(grads),
            false_fn=lambda: tf.constant(1.0)))

    new_global_step = tf.cond(pred=all_are_finite, true_fn=lambda: global_step + 1, false_fn=lambda: global_step)
    new_global_step = tf.identity(new_global_step, name='step_update')

    with tf.control_dependencies([global_step.assign(new_global_step)]):
      train_op = optimizer.apply_gradients(
          list(zip(clipped_grads, tvars)), global_step=global_step)
  return train_op


class AdamWeightDecayOptimizer(tf.compat.v1.train.Optimizer):
  """A basic Adam optimizer that includes "correct" L2 weight decay."""

  def __init__(self,
               learning_rate,
               weight_decay_rate=0.0,
               beta_1=0.9,
               beta_2=0.999,
               epsilon=1e-6,
               exclude_from_weight_decay=None,
               name="AdamWeightDecayOptimizer"):
    """Constructs a AdamWeightDecayOptimizer."""
    super(AdamWeightDecayOptimizer, self).__init__(False, name)

    self.learning_rate = tf.identity(learning_rate, name='learning_rate')
    self.weight_decay_rate = weight_decay_rate
    self.beta_1 = beta_1
    self.beta_2 = beta_2
    self.epsilon = epsilon
    self.exclude_from_weight_decay = exclude_from_weight_decay

  def apply_gradients(self, grads_and_vars, global_step=None, name=None,
                      manual_fp16=False):
    """See base class."""
    assignments = []
    for (grad, param) in grads_and_vars:
      if grad is None or param is None:
        continue

      param_name = self._get_variable_name(param.name)
      has_shadow = manual_fp16 and param.dtype.base_dtype != tf.float32
      if has_shadow:
        # create shadow fp32 weights for fp16 variable
        param_fp32 = tf.compat.v1.get_variable(
            name=param_name + "/shadow",
            dtype=tf.float32,
            trainable=False,
            initializer=tf.cast(param.initialized_value(), tf.float32))
      else:
        param_fp32 = param

      m = tf.compat.v1.get_variable(
          name=param_name + "/adam_m",
          shape=param.shape.as_list(),
          dtype=tf.float32,
          trainable=False,
          initializer=tf.compat.v1.zeros_initializer())
      v = tf.compat.v1.get_variable(
          name=param_name + "/adam_v",
          shape=param.shape.as_list(),
          dtype=tf.float32,
          trainable=False,
          initializer=tf.compat.v1.zeros_initializer())

      # Standard Adam update.
      next_m = (
          tf.multiply(self.beta_1, m) + tf.multiply(1.0 - self.beta_1, grad))
      next_v = (
          tf.multiply(self.beta_2, v) + tf.multiply(1.0 - self.beta_2,
                                                    tf.square(grad)))

      update = next_m * tf.math.rsqrt(next_v + self.epsilon * self.epsilon)

      # Just adding the square of the weights to the loss function is *not*
      # the correct way of using L2 regularization/weight decay with Adam,
      # since that will interact with the m and v parameters in strange ways.
      #
      # Instead we want to decay the weights in a manner that doesn't interact
      # with the m/v parameters. This is equivalent to adding the square
      # of the weights to the loss with plain (non-momentum) SGD.
      if self._do_use_weight_decay(param_name):
        update += self.weight_decay_rate * param_fp32

      update_with_lr = self.learning_rate * update

      next_param = param_fp32 - update_with_lr

      if has_shadow:
        # cast shadow fp32 weights to fp16 and assign to trainable variable
        param.assign(tf.cast(next_param, param.dtype.base_dtype))
      assignments.extend(
          [param_fp32.assign(next_param),
           m.assign(next_m),
           v.assign(next_v)])
    return tf.group(*assignments, name=name)

  def _do_use_weight_decay(self, param_name):
    """Whether to use L2 weight decay for `param_name`."""
    if not self.weight_decay_rate:
      return False
    if self.exclude_from_weight_decay:
      for r in self.exclude_from_weight_decay:
        if re.search(r, param_name) is not None:
          return False
    return True

  def _get_variable_name(self, param_name):
    """Get the variable name from the tensor name."""
    m = re.match("^(.*):\\d+$", param_name)
    if m is not None:
      param_name = m.group(1)
    return param_name

  # This code originally was a WA for this issue:
  #    See: https://jira.habana-labs.com/browse/SW-19371
  # However, the root issue has been fixed and is no longer required.
  #
  # It turned out that this function needs to be uncommented to speed up the BERT finetuning training.
  #    See: https://jira.habana-labs.com/browse/SW-19126
  #
  # At this moment, enabling SAO leads to an immediate crash:
  #    See: https://jira.habana-labs.com/browse/SW-19688
  #
  def compute_gradients(self, loss, var_list=None,
                        gate_gradients=tf.compat.v1.train.Optimizer.GATE_OP,
                        aggregation_method=None,
                        colocate_gradients_with_ops=False,
                        grad_loss=None):
    assert gate_gradients == tf.compat.v1.train.Optimizer.GATE_OP
    assert aggregation_method is None
    assert not colocate_gradients_with_ops
    assert grad_loss is None

    grads = tf.gradients(ys=loss, xs=var_list)
    grads_and_vars = list(zip(grads, var_list))
    return grads_and_vars


class LAMBOptimizer(tf.compat.v1.train.Optimizer):
  """A LAMB optimizer that includes "correct" L2 weight decay."""

  def __init__(self,
               learning_rate,
               weight_decay_rate=0.0,
               beta_1=0.9,
               beta_2=0.999,
               epsilon=1e-6,
               exclude_from_weight_decay=None,
               name="LAMBOptimizer"):
    """Constructs a LAMBOptimizer."""
    super(LAMBOptimizer, self).__init__(False, name)

    self.learning_rate = tf.identity(learning_rate, name='learning_rate')
    self.weight_decay_rate = weight_decay_rate
    self.beta_1 = beta_1
    self.beta_2 = beta_2
    self.epsilon = epsilon
    self.exclude_from_weight_decay = exclude_from_weight_decay

  def apply_gradients(self, grads_and_vars, global_step, name=None,
                      manual_fp16=False):
    """See base class."""
    assignments = []
    steps = tf.cast(global_step, tf.float32)
    for (grad, param) in grads_and_vars:
      if grad is None or param is None:
        continue

      param_name = self._get_variable_name(param.name)
      has_shadow = manual_fp16 and param.dtype.base_dtype != tf.float32
      if has_shadow:
        # create shadow fp32 weights for fp16 variable
        param_fp32 = tf.compat.v1.get_variable(
            name=param_name + "/shadow",
            dtype=tf.float32,
            trainable=False,
            initializer=tf.cast(param.initialized_value(), tf.float32))
      else:
        param_fp32 = param

      m = tf.compat.v1.get_variable(
          name=param_name + "/adam_m",
          shape=param.shape.as_list(),
          dtype=tf.float32,
          trainable=False,
          initializer=tf.compat.v1.zeros_initializer())
      v = tf.compat.v1.get_variable(
          name=param_name + "/adam_v",
          shape=param.shape.as_list(),
          dtype=tf.float32,
          trainable=False,
          initializer=tf.compat.v1.zeros_initializer())

      # LAMB update
      next_m = (
          tf.multiply(self.beta_1, m) + tf.multiply(1.0 - self.beta_1, grad))
      next_v = (
          tf.multiply(self.beta_2, v) + tf.multiply(1.0 - self.beta_2,
                                                    tf.square(grad)))

      beta1_correction = (1 - self.beta_1 ** steps)
      beta2_correction = (1 - self.beta_2 ** steps)

      next_m_unbiased = next_m / beta1_correction
      next_v_unbiased = next_v / beta2_correction

      update = next_m_unbiased / (tf.sqrt(next_v_unbiased) + self.epsilon)

      # Just adding the square of the weights to the loss function is *not*
      # the correct way of using L2 regularization/weight decay with Adam,
      # since that will interact with the m and v parameters in strange ways.
      #
      # Instead we want to decay the weights in a manner that doesn't interact
      # with the m/v parameters. This is equivalent to adding the square
      # of the weights to the loss with plain (non-momentum) SGD.
      if self._do_use_weight_decay(param_name):
        update += self.weight_decay_rate * param_fp32

      w_norm = linalg_ops.norm(param, ord=2)
      g_norm = linalg_ops.norm(update, ord=2)
      ratio = array_ops.where(math_ops.greater(w_norm, 0), array_ops.where(
          math_ops.greater(g_norm, 0), (w_norm / g_norm), 1.0), 1.0)

      update_with_lr = ratio * self.learning_rate * update

      next_param = param_fp32 - update_with_lr

      if has_shadow:
        # cast shadow fp32 weights to fp16 and assign to trainable variable
        param.assign(tf.cast(next_param, param.dtype.base_dtype))
      assignments.extend(
          [param_fp32.assign(next_param),
           m.assign(next_m),
           v.assign(next_v)])
    return tf.group(*assignments, name=name)

  def _do_use_weight_decay(self, param_name):
    """Whether to use L2 weight decay for `param_name`."""
    if not self.weight_decay_rate:
      return False
    if self.exclude_from_weight_decay:
      for r in self.exclude_from_weight_decay:
        if re.search(r, param_name) is not None:
          return False
    return True

  def _get_variable_name(self, param_name):
    """Get the variable name from the tensor name."""
    m = re.match("^(.*):\\d+$", param_name)
    if m is not None:
      param_name = m.group(1)
    return param_name