File size: 10,937 Bytes
b100e1c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2022 The T5X 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.

"""Loss functions."""
import enum
from typing import Tuple, Mapping, Optional, Union

from flax.training import common_utils
import jax
import jax.numpy as jnp
import numpy as np


@jax.custom_vjp
def cross_entropy_with_logits(logits: jnp.ndarray, targets: jnp.ndarray,
                              z_loss: float) -> jnp.ndarray:
  """Computes cross entropy loss with stable custom gradient.

  Computes a stabilized-gradient version of:
    -jnp.sum(targets * nn.log_softmax(logits), axis=-1)

  If z_loss > 0, then an auxiliary loss equal to z_loss*log(z)^2
  will be added to the cross entropy loss (z = softmax normalization constant).
  The two uses of z_loss are:
  1. To keep the logits from drifting too far from zero, which can cause
     unacceptable roundoff errors in bfloat16.
  2. To encourage the logits to be normalized log-probabilities.

  Args:
    logits: [batch, length, num_classes] float array.
    targets: categorical one-hot targets [batch, length, num_classes] float
      array.
    z_loss: coefficient for auxilliary z-loss loss term.

  Returns:
    tuple with the total loss and the z_loss, both
    float arrays with shape [batch, length].
  """
  logits_sum = jax.scipy.special.logsumexp(logits, axis=-1, keepdims=True)
  log_softmax = logits - logits_sum
  loss = -jnp.sum(targets * log_softmax, axis=-1)
  # Add auxilliary z-loss term.
  log_z = jnp.squeeze(logits_sum, axis=-1)
  total_z_loss = z_loss * jax.lax.square(log_z)
  loss += total_z_loss
  return loss, total_z_loss


def _cross_entropy_with_logits_fwd(
    logits: jnp.ndarray,
    targets: jnp.ndarray,
    z_loss: float = 0.0
) -> Tuple[jnp.ndarray, Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray, jnp
                              .ndarray, jnp.ndarray, jnp.ndarray, jnp.ndarray]]:
  """Forward-mode of `cross_entropy_with_logits`."""
  max_logit = logits.max(axis=-1, keepdims=True)
  shifted = logits - max_logit
  exp_shifted = jnp.exp(shifted)
  sum_exp = jnp.sum(exp_shifted, axis=-1, keepdims=True)
  log_softmax = shifted - jnp.log(sum_exp)
  loss = -jnp.sum(targets * log_softmax, axis=-1)
  # Add auxilliary z-loss term.
  log_z = jnp.squeeze(jnp.log(sum_exp) + max_logit, axis=-1)
  total_z_loss = z_loss * jax.lax.square(log_z)
  loss += total_z_loss
  return (loss, total_z_loss), (logits, targets, z_loss, exp_shifted, sum_exp,
                                log_softmax, log_z)


def _cross_entropy_with_logits_bwd(
    res: Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray, jnp.ndarray, jnp.ndarray,
               jnp.ndarray, jnp.ndarray], g: Tuple[jnp.ndarray, jnp.ndarray]
) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
  """Backward-mode of `cross_entropy_with_logits`."""
  g = g[0]  # Ignore z_loss component as that is only used for logging.
  logits, targets, z_loss, exp_shifted, sum_exp, log_softmax, log_z = res
  # z-loss term adds the (2 * z_loss * log_z) factor.
  deriv = (
      jnp.expand_dims(1 + 2 * z_loss * log_z, -1) * exp_shifted / sum_exp -
      targets)
  g_logits = jnp.expand_dims(g, axis=-1) * deriv
  g_targets = -jnp.expand_dims(g, axis=-1) * log_softmax
  return (jnp.asarray(g_logits,
                      logits.dtype), jnp.asarray(g_targets, targets.dtype),
          jnp.array(0.0))  # sets z-loss coeff gradient to 0


cross_entropy_with_logits.defvjp(_cross_entropy_with_logits_fwd,
                                 _cross_entropy_with_logits_bwd)


def compute_weighted_cross_entropy(
    logits: jnp.ndarray,
    targets: jnp.ndarray,
    weights: Optional[jnp.ndarray] = None,
    label_smoothing: float = 0.0,
    z_loss: float = 0.0,
    loss_normalizing_factor: Optional[float] = None
) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
  """Compute weighted cross entropy and entropy for log probs and targets.

  Args:
   logits: [batch, length, num_classes] float array.
   targets: categorical targets [batch, length] int array.
   weights: None or array of shape [batch, length].
   label_smoothing: label smoothing constant, used to determine the on and off
     values.
   z_loss: coefficient for auxiliary z-loss loss term.
   loss_normalizing_factor: Constant to divide loss by. If not specified, loss
     will not be normalized. Intended for backward compatibility with T5-MTF
     training. Should not normally be used.

  Returns:
    Tuple of scalar loss, z_loss, and weight sum.
  """
  if logits.ndim != targets.ndim + 1:
    raise ValueError('Incorrect shapes. Got shape %s logits and %s targets' %
                     (str(logits.shape), str(targets.shape)))
  vocab_size = logits.shape[-1]
  confidence = 1.0 - label_smoothing
  low_confidence = (1.0 - confidence) / (vocab_size - 1)
  normalizing_constant = -(
      confidence * jnp.log(confidence) +
      (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20))
  soft_targets = common_utils.onehot(
      targets, vocab_size, on_value=confidence, off_value=low_confidence)
  total_loss, total_z_loss = cross_entropy_with_logits(
      logits, soft_targets, z_loss=z_loss)
  total_loss = total_loss - normalizing_constant

  weight_sum = np.prod(targets.shape)
  if weights is not None:
    total_loss = total_loss * weights
    total_z_loss = total_z_loss * weights
    weight_sum = jnp.sum(weights)

  # By default, we do not normalize loss based on anything.
  # We don't normalize based on batch size because the optimizers we use are
  # pretty much scale invariant, so this simplifies things.
  # We don't normalize based on number of non-padding tokens in order to treat
  # each token as equally important regardless of sequence length.
  if loss_normalizing_factor is not None:
    total_loss /= loss_normalizing_factor
    total_z_loss /= loss_normalizing_factor
  return jnp.sum(total_loss), jnp.sum(total_z_loss), weight_sum


@enum.unique
class SpecialLossNormalizingFactor(enum.Enum):
  """Specially calcualted loss_normalizing_factors, that are not a constant.

  Attributes:
    NUM_REAL_TARGET_TOKENS: Whether to divide the loss by the number of real
      (non-padding) tokens in the current target batch. If
      'decoder_loss_weights' are specified, it will be the sum of the weights.
      Otherwise it will be the number of non-zero 'decoder_target_tokens'.
    NUM_TOTAL_TARGET_TOKENS: Whether to divide the loss by the total number of
      target tokens, i.e., batch_size * target_seq_length (including padding).
    AVERAGE_PER_SEQUENCE: This will first compute the per-sequence loss
      (averaged over the number of real target tokens in the sequence), and then
      compute the average of that over the sequences. This can be preferable to
      NUM_REAL_TARGET_TOKENS for finetuning, because it will weigh all examples
      equally, regardless of sequence length (which can be especially important
      for multi-task finetuning).
  """
  NUM_REAL_TARGET_TOKENS = 1
  NUM_TOTAL_TARGET_TOKENS = 2
  AVERAGE_PER_SEQUENCE = 3


def convert_special_loss_normalizing_factor_to_enum(
    x: str) -> SpecialLossNormalizingFactor:
  """Converts stringified version of LNF to an enum.

  This is useful because gin dynamic registration does not (currently)
  have support for enum.

  Args:
    x: stringified version of SpecialLossNormalizingFactor enum.

  Returns:
    SpecialLossNormalizingFactor enum instance.
  """
  x = x.upper()
  if x == 'NUM_REAL_TARGET_TOKENS':
    return SpecialLossNormalizingFactor.NUM_REAL_TARGET_TOKENS
  if x == 'NUM_TOTAL_TARGET_TOKENS':
    return SpecialLossNormalizingFactor.NUM_TOTAL_TARGET_TOKENS
  if x == 'AVERAGE_PER_SEQUENCE':
    return SpecialLossNormalizingFactor.AVERAGE_PER_SEQUENCE
  raise ValueError(
      'Could not convert string \"%s\" to SpecialLossNormalizingFactor' % x)


def get_loss_normalizing_factor_and_weights(
    loss_normalizing_factor: Optional[Union[float, int, str,
                                            SpecialLossNormalizingFactor]],
    batch: Mapping[str, jnp.ndarray]):
  """Get the float loss_normalizing_factor and loss weights.

  If loss_normalizing_factor is float or None, this will simply return the
  input loss_normalizing_factor and batch.

  If loss_normalizing_factor is a SpecialLossNormalizingFactor, it will
  return a float loss_normalizing_factor and loss weights corresponding to
  the special LNF. See SpecialLossNormalizingFactor for more details.

  Args:
    loss_normalizing_factor: The input LNF, which may be a float, None, or
      SpecialLossNormalizingFactor (or a stringified SLNF).
    batch: Input data batch.

  Returns:
    Tuple of (output_loss_normalizing_factor, loss_weights).
      'output_loss_normalizing_factor' is a scalar float (Python float
      or jnp float).
      'loss_weights' is the per token loss weight JNP array.
  """

  loss_weights = batch.get('decoder_loss_weights', None)
  if (loss_normalizing_factor is None or
      not isinstance(loss_normalizing_factor,
                     (str, SpecialLossNormalizingFactor))):
    return (loss_normalizing_factor, loss_weights)

  if isinstance(loss_normalizing_factor, str):
    loss_normalizing_factor = convert_special_loss_normalizing_factor_to_enum(
        loss_normalizing_factor)

  # If `loss_weights` are not provided, we assume that the padding id is 0 and
  # that non-padding tokens in the decoder all correspond to the positions
  # where loss should be taken. If more fine-grained behavior (e.g., taking
  # loss on subset of 'decoder_target_tokens') is desired, provide
  # `loss_weights` that account for this.
  if loss_weights is None:
    loss_weights = jnp.asarray(batch['decoder_target_tokens'] > 0, jnp.float32)

  output_normalizing_factor = None
  if (loss_normalizing_factor ==
      SpecialLossNormalizingFactor.NUM_REAL_TARGET_TOKENS):
    output_normalizing_factor = jnp.sum(loss_weights)
  elif (loss_normalizing_factor ==
        SpecialLossNormalizingFactor.NUM_TOTAL_TARGET_TOKENS):
    output_normalizing_factor = np.prod(batch['decoder_target_tokens'].shape)
  elif (loss_normalizing_factor ==
        SpecialLossNormalizingFactor.AVERAGE_PER_SEQUENCE):
    loss_weights /= jnp.sum(loss_weights, axis=-1, keepdims=True) + 1e-3
    output_normalizing_factor = jnp.sum(loss_weights)
  else:
    raise ValueError('Unsupported value of loss_normalizing_factor: %s' %
                     str(loss_normalizing_factor))

  return (output_normalizing_factor, loss_weights)