/
nat_loss.py
173 lines (145 loc) · 6.09 KB
/
nat_loss.py
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
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import math
import torch.nn.functional as F
import torch
from torch import Tensor
from fairseq import metrics, utils
from fairseq.criterions import FairseqCriterion, register_criterion
@register_criterion("nat_loss")
class LabelSmoothedDualImitationCriterion(FairseqCriterion):
def __init__(self, task, label_smoothing):
super().__init__(task)
self.label_smoothing = label_smoothing
@staticmethod
def add_args(parser):
"""Add criterion-specific arguments to the parser."""
parser.add_argument(
'--label-smoothing',
default=0.,
type=float,
metavar='D',
help='epsilon for label smoothing, 0 means no label smoothing',
)
def _compute_loss(
self, outputs, targets, masks=None, label_smoothing=0.0, name="loss", factor=1.0
):
"""
outputs: batch x len x d_model
targets: batch x len
masks: batch x len
policy_logprob: if there is some policy
depends on the likelihood score as rewards.
"""
def mean_ds(x: Tensor, dim=None) -> Tensor:
return (
x.float().mean().type_as(x)
if dim is None
else x.float().mean(dim).type_as(x)
)
if masks is not None:
outputs, targets = outputs[masks], targets[masks]
if masks is not None and not masks.any():
nll_loss = torch.tensor(0)
loss = nll_loss
else:
logits = F.log_softmax(outputs, dim=-1)
if targets.dim() == 1:
losses = F.nll_loss(logits, targets.to(logits.device), reduction='none')
else: # soft-labels
losses = F.kl_div(logits, targets.to(logits.device), reduction='none')
losses = losses.sum(-1)
nll_loss = mean_ds(losses)
if label_smoothing > 0:
loss = nll_loss * (
1 - label_smoothing) - mean_ds(logits) * label_smoothing
else:
loss = nll_loss
loss = loss * factor
return {"name": name, "loss": loss, "nll_loss": nll_loss, "factor": factor}
def _custom_loss(self, loss, name="loss", factor=1.0):
return {"name": name, "loss": loss, "factor": factor}
def forward(self, model, sample, reduce=True):
"""Compute the loss for the given sample.
Returns a tuple with three elements:
1) the loss
2) the sample size, which is used as the denominator for the gradient
3) logging outputs to display while training
"""
nsentences, ntokens = sample["nsentences"], sample["ntokens"]
# B x T
src_tokens, src_lengths = (
sample["net_input"]["src_tokens"],
sample["net_input"]["src_lengths"],
)
tgt_tokens, prev_output_tokens = sample["target"], sample["prev_target"]
outputs = model(src_tokens, src_lengths, prev_output_tokens, tgt_tokens)
losses, nll_loss = [], []
for obj in outputs:
if outputs[obj].get("loss", None) is None:
_losses = self._compute_loss(
outputs[obj].get("out"),
outputs[obj].get("tgt"),
outputs[obj].get("mask", None),
outputs[obj].get("ls", 0.0),
name=obj + '-loss',
factor=outputs[obj].get("factor", 1.0)
)
else:
_losses = self._custom_loss(
outputs[obj].get("loss"),
name=obj + '-loss',
factor=outputs[obj].get("factor", 1.0)
)
losses += [_losses]
if outputs[obj].get("nll_loss", False):
nll_loss += [_losses.get("nll_loss", 0.0)]
loss = sum(l["loss"] for l in losses)
nll_loss = sum(l for l in nll_loss) if len(nll_loss) > 0 \
else loss.new_tensor(0)
# NOTE:
# we don't need to use sample_size as denominator for the gradient
# here sample_size is just used for logging
sample_size = 1
logging_output = {
"loss": loss.data,
"nll_loss": nll_loss.data,
"ntokens": ntokens,
"nsentences": nsentences,
"sample_size": sample_size,
}
for l in losses:
logging_output[l["name"]] = (
utils.item(l["loss"].data / l["factor"])
if reduce
else l[["loss"]].data / l["factor"]
)
return loss, sample_size, logging_output
@staticmethod
def reduce_metrics(logging_outputs) -> None:
"""Aggregate logging outputs from data parallel training."""
sample_size = utils.item(sum(log.get("sample_size", 0) for log in logging_outputs))
loss = utils.item(sum(log.get("loss", 0) for log in logging_outputs))
nll_loss = utils.item(sum(log.get("nll_loss", 0) for log in logging_outputs))
metrics.log_scalar('loss', loss / sample_size / math.log(2), sample_size, round=3)
metrics.log_scalar('nll_loss', nll_loss / sample_size / math.log(2), sample_size, round=3)
metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['loss'].avg))
for key in logging_outputs[0]:
if key[-5:] == "-loss":
val = sum(log.get(key, 0) for log in logging_outputs)
metrics.log_scalar(
key[:-5],
val / sample_size / math.log(2) if sample_size > 0 else 0.0,
sample_size,
round=3,
)
@staticmethod
def logging_outputs_can_be_summed() -> bool:
"""
Whether the logging outputs returned by `forward` can be summed
across workers prior to calling `reduce_metrics`. Setting this
to True will improves distributed training speed.
"""
return True