File size: 11,065 Bytes
d5175d3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# All rights reserved.
#
# This source code is licensed under the license found in the LICENSE file in
# the root directory of this source tree. An additional grant of patent rights
# can be found in the PATENTS file in the same directory.

import math
from argparse import Namespace
from dataclasses import dataclass, field
from omegaconf import II
from typing import Optional

import torch
import torch.nn.functional as F
from fairseq import metrics, utils
from fairseq.criterions import FairseqCriterion, register_criterion
from fairseq.dataclass import FairseqDataclass
from fairseq.data.data_utils import post_process
from fairseq.tasks import FairseqTask
from fairseq.logging.meters import safe_round


@dataclass
class CtcCriterionConfig(FairseqDataclass):
    zero_infinity: bool = field(
        default=False,
        metadata={"help": "zero inf loss when source length <= target length"},
    )
    sentence_avg: bool = II("optimization.sentence_avg")
    post_process: str = field(
        default="letter",
        metadata={
            "help": "how to post process predictions into words. can be letter, "
            "wordpiece, BPE symbols, etc. "
            "See fairseq.data.data_utils.post_process() for full list of options"
        },
    )
    wer_kenlm_model: Optional[str] = field(
        default=None,
        metadata={
            "help": "if this is provided, use kenlm to compute wer (along with other wer_* args)"
        },
    )
    wer_lexicon: Optional[str] = field(
        default=None,
        metadata={"help": "lexicon to use with wer_kenlm_model"},
    )
    wer_lm_weight: float = field(
        default=2.0,
        metadata={"help": "lm weight to use with wer_kenlm_model"},
    )
    wer_word_score: float = field(
        default=-1.0,
        metadata={"help": "lm word score to use with wer_kenlm_model"},
    )

    wer_args: Optional[str] = field(
        default=None,
        metadata={
            "help": "DEPRECATED: tuple of (wer_kenlm_model, wer_lexicon, wer_lm_weight, wer_word_score)"
        },
    )


@register_criterion("ctc", dataclass=CtcCriterionConfig)
class CtcCriterion(FairseqCriterion):
    def __init__(self, cfg: CtcCriterionConfig, task: FairseqTask):
        super().__init__(task)
        self.blank_idx = (
            task.target_dictionary.index(task.blank_symbol)
            if hasattr(task, "blank_symbol")
            else 0
        )
        self.pad_idx = task.target_dictionary.pad()
        self.eos_idx = task.target_dictionary.eos()
        self.post_process = cfg.post_process

        if cfg.wer_args is not None:
            (
                cfg.wer_kenlm_model,
                cfg.wer_lexicon,
                cfg.wer_lm_weight,
                cfg.wer_word_score,
            ) = eval(cfg.wer_args)

        if cfg.wer_kenlm_model is not None:
            from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder

            dec_args = Namespace()
            dec_args.nbest = 1
            dec_args.criterion = "ctc"
            dec_args.kenlm_model = cfg.wer_kenlm_model
            dec_args.lexicon = cfg.wer_lexicon
            dec_args.beam = 50
            dec_args.beam_size_token = min(50, len(task.target_dictionary))
            dec_args.beam_threshold = min(50, len(task.target_dictionary))
            dec_args.lm_weight = cfg.wer_lm_weight
            dec_args.word_score = cfg.wer_word_score
            dec_args.unk_weight = -math.inf
            dec_args.sil_weight = 0

            self.w2l_decoder = W2lKenLMDecoder(dec_args, task.target_dictionary)
        else:
            self.w2l_decoder = None

        self.zero_infinity = cfg.zero_infinity
        self.sentence_avg = cfg.sentence_avg

    def forward(self, model, sample, reduce=True):
        net_output = model(**sample["net_input"])
        lprobs = model.get_normalized_probs(
            net_output, log_probs=True
        ).contiguous()  # (T, B, C) from the encoder

        if "src_lengths" in sample["net_input"]:
            input_lengths = sample["net_input"]["src_lengths"]
        else:
            if net_output["padding_mask"] is not None:
                non_padding_mask = ~net_output["padding_mask"]
                input_lengths = non_padding_mask.long().sum(-1)
            else:
                input_lengths = lprobs.new_full(
                    (lprobs.size(1),), lprobs.size(0), dtype=torch.long
                )

        pad_mask = (sample["target"] != self.pad_idx) & (
            sample["target"] != self.eos_idx
        )
        targets_flat = sample["target"].masked_select(pad_mask)
        if "target_lengths" in sample:
            target_lengths = sample["target_lengths"]
        else:
            target_lengths = pad_mask.sum(-1)

        with torch.backends.cudnn.flags(enabled=False):
            loss = F.ctc_loss(
                lprobs,
                targets_flat,
                input_lengths,
                target_lengths,
                blank=self.blank_idx,
                reduction="sum",
                zero_infinity=self.zero_infinity,
            )

        ntokens = (
            sample["ntokens"] if "ntokens" in sample else target_lengths.sum().item()
        )

        sample_size = sample["target"].size(0) if self.sentence_avg else ntokens
        logging_output = {
            "loss": utils.item(loss.data),  # * sample['ntokens'],
            "ntokens": ntokens,
            "nsentences": sample["id"].numel(),
            "sample_size": sample_size,
        }

        if not model.training:
            import editdistance

            with torch.no_grad():
                lprobs_t = lprobs.transpose(0, 1).float().contiguous().cpu()

                c_err = 0
                c_len = 0
                w_errs = 0
                w_len = 0
                wv_errs = 0
                for lp, t, inp_l in zip(
                    lprobs_t,
                    sample["target_label"]
                    if "target_label" in sample
                    else sample["target"],
                    input_lengths,
                ):
                    lp = lp[:inp_l].unsqueeze(0)

                    decoded = None
                    if self.w2l_decoder is not None:
                        decoded = self.w2l_decoder.decode(lp)
                        if len(decoded) < 1:
                            decoded = None
                        else:
                            decoded = decoded[0]
                            if len(decoded) < 1:
                                decoded = None
                            else:
                                decoded = decoded[0]

                    p = (t != self.task.target_dictionary.pad()) & (
                        t != self.task.target_dictionary.eos()
                    )
                    targ = t[p]
                    targ_units = self.task.target_dictionary.string(targ)
                    targ_units_arr = targ.tolist()

                    toks = lp.argmax(dim=-1).unique_consecutive()
                    pred_units_arr = toks[toks != self.blank_idx].tolist()

                    c_err += editdistance.eval(pred_units_arr, targ_units_arr)
                    c_len += len(targ_units_arr)

                    targ_words = post_process(targ_units, self.post_process).split()

                    pred_units = self.task.target_dictionary.string(pred_units_arr)
                    pred_words_raw = post_process(pred_units, self.post_process).split()

                    if decoded is not None and "words" in decoded:
                        pred_words = decoded["words"]
                        w_errs += editdistance.eval(pred_words, targ_words)
                        wv_errs += editdistance.eval(pred_words_raw, targ_words)
                    else:
                        dist = editdistance.eval(pred_words_raw, targ_words)
                        w_errs += dist
                        wv_errs += dist

                    w_len += len(targ_words)

                logging_output["wv_errors"] = wv_errs
                logging_output["w_errors"] = w_errs
                logging_output["w_total"] = w_len
                logging_output["c_errors"] = c_err
                logging_output["c_total"] = c_len

        return loss, sample_size, logging_output

    @staticmethod
    def reduce_metrics(logging_outputs) -> None:
        """Aggregate logging outputs from data parallel training."""

        loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs))
        ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs))
        nsentences = utils.item(
            sum(log.get("nsentences", 0) for log in logging_outputs)
        )
        sample_size = utils.item(
            sum(log.get("sample_size", 0) for log in logging_outputs)
        )

        metrics.log_scalar(
            "loss", loss_sum / sample_size / math.log(2), sample_size, round=3
        )
        metrics.log_scalar("ntokens", ntokens)
        metrics.log_scalar("nsentences", nsentences)
        if sample_size != ntokens:
            metrics.log_scalar(
                "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3
            )

        c_errors = sum(log.get("c_errors", 0) for log in logging_outputs)
        metrics.log_scalar("_c_errors", c_errors)
        c_total = sum(log.get("c_total", 0) for log in logging_outputs)
        metrics.log_scalar("_c_total", c_total)
        w_errors = sum(log.get("w_errors", 0) for log in logging_outputs)
        metrics.log_scalar("_w_errors", w_errors)
        wv_errors = sum(log.get("wv_errors", 0) for log in logging_outputs)
        metrics.log_scalar("_wv_errors", wv_errors)
        w_total = sum(log.get("w_total", 0) for log in logging_outputs)
        metrics.log_scalar("_w_total", w_total)

        if c_total > 0:
            metrics.log_derived(
                "uer",
                lambda meters: safe_round(
                    meters["_c_errors"].sum * 100.0 / meters["_c_total"].sum, 3
                )
                if meters["_c_total"].sum > 0
                else float("nan"),
            )
        if w_total > 0:
            metrics.log_derived(
                "wer",
                lambda meters: safe_round(
                    meters["_w_errors"].sum * 100.0 / meters["_w_total"].sum, 3
                )
                if meters["_w_total"].sum > 0
                else float("nan"),
            )
            metrics.log_derived(
                "raw_wer",
                lambda meters: safe_round(
                    meters["_wv_errors"].sum * 100.0 / meters["_w_total"].sum, 3
                )
                if meters["_w_total"].sum > 0
                else float("nan"),
            )

    @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