From 3ff62dbb97684837f76c75f2defe6d0c77bc6d48 Mon Sep 17 00:00:00 2001
From: speech_asr <wangjiaming.wjm@alibaba-inc.com>
Date: 星期一, 13 三月 2023 16:04:27 +0800
Subject: [PATCH] update ola

---
 funasr/models/e2e_diar_eend_ola.py |  395 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 395 insertions(+), 0 deletions(-)

diff --git a/funasr/models/e2e_diar_eend_ola.py b/funasr/models/e2e_diar_eend_ola.py
new file mode 100644
index 0000000..967c0d4
--- /dev/null
+++ b/funasr/models/e2e_diar_eend_ola.py
@@ -0,0 +1,395 @@
+# Copyright ESPnet (https://github.com/espnet/espnet). All Rights Reserved.
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+import logging
+import torch
+from contextlib import contextmanager
+from distutils.version import LooseVersion
+from funasr.layers.abs_normalize import AbsNormalize
+from funasr.losses.label_smoothing_loss import (
+    LabelSmoothingLoss,  # noqa: H301
+)
+from funasr.models.ctc import CTC
+from funasr.models.decoder.abs_decoder import AbsDecoder
+from funasr.models.encoder.abs_encoder import AbsEncoder
+from funasr.models.frontend.abs_frontend import AbsFrontend
+from funasr.models.postencoder.abs_postencoder import AbsPostEncoder
+from funasr.models.preencoder.abs_preencoder import AbsPreEncoder
+from funasr.models.specaug.abs_specaug import AbsSpecAug
+from funasr.modules.add_sos_eos import add_sos_eos
+from funasr.modules.e2e_asr_common import ErrorCalculator
+from funasr.modules.eend_ola.encoder import TransformerEncoder
+from funasr.modules.eend_ola.encoder_decoder_attractor import EncoderDecoderAttractor
+from funasr.modules.eend_ola.utils.power import generate_mapping_dict
+from funasr.modules.nets_utils import th_accuracy
+from funasr.torch_utils.device_funcs import force_gatherable
+from funasr.train.abs_espnet_model import AbsESPnetModel
+from typeguard import check_argument_types
+from typing import Dict
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):
+    from torch.cuda.amp import autocast
+else:
+    # Nothing to do if torch<1.6.0
+    @contextmanager
+    def autocast(enabled=True):
+        yield
+
+
+class DiarEENDOLAModel(AbsESPnetModel):
+    """CTC-attention hybrid Encoder-Decoder model"""
+
+    def __init__(
+            self,
+            encoder: TransformerEncoder,
+            eda: EncoderDecoderAttractor,
+            max_n_speaker: int = 8,
+            attractor_loss_weight: float = 1.0,
+            mapping_dict=None,
+            **kwargs,
+    ):
+        assert check_argument_types()
+
+        super().__init__()
+        self.encoder = encoder
+        self.eda = eda
+        self.attractor_loss_weight = attractor_loss_weight
+        self.max_n_speaker = max_n_speaker
+        if mapping_dict is None:
+            mapping_dict = generate_mapping_dict(max_speaker_num=self.max_n_speaker)
+            self.mapping_dict = mapping_dict
+
+    def forward(
+            self,
+            speech: torch.Tensor,
+            speech_lengths: torch.Tensor,
+            text: torch.Tensor,
+            text_lengths: torch.Tensor,
+    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]:
+        """Frontend + Encoder + Decoder + Calc loss
+
+        Args:
+            speech: (Batch, Length, ...)
+            speech_lengths: (Batch, )
+            text: (Batch, Length)
+            text_lengths: (Batch,)
+        """
+        assert text_lengths.dim() == 1, text_lengths.shape
+        # Check that batch_size is unified
+        assert (
+                speech.shape[0]
+                == speech_lengths.shape[0]
+                == text.shape[0]
+                == text_lengths.shape[0]
+        ), (speech.shape, speech_lengths.shape, text.shape, text_lengths.shape)
+        batch_size = speech.shape[0]
+
+        # for data-parallel
+        text = text[:, : text_lengths.max()]
+
+        # 1. Encoder
+        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
+        intermediate_outs = None
+        if isinstance(encoder_out, tuple):
+            intermediate_outs = encoder_out[1]
+            encoder_out = encoder_out[0]
+
+        loss_att, acc_att, cer_att, wer_att = None, None, None, None
+        loss_ctc, cer_ctc = None, None
+        stats = dict()
+
+        # 1. CTC branch
+        if self.ctc_weight != 0.0:
+            loss_ctc, cer_ctc = self._calc_ctc_loss(
+                encoder_out, encoder_out_lens, text, text_lengths
+            )
+
+            # Collect CTC branch stats
+            stats["loss_ctc"] = loss_ctc.detach() if loss_ctc is not None else None
+            stats["cer_ctc"] = cer_ctc
+
+        # Intermediate CTC (optional)
+        loss_interctc = 0.0
+        if self.interctc_weight != 0.0 and intermediate_outs is not None:
+            for layer_idx, intermediate_out in intermediate_outs:
+                # we assume intermediate_out has the same length & padding
+                # as those of encoder_out
+                loss_ic, cer_ic = self._calc_ctc_loss(
+                    intermediate_out, encoder_out_lens, text, text_lengths
+                )
+                loss_interctc = loss_interctc + loss_ic
+
+                # Collect Intermedaite CTC stats
+                stats["loss_interctc_layer{}".format(layer_idx)] = (
+                    loss_ic.detach() if loss_ic is not None else None
+                )
+                stats["cer_interctc_layer{}".format(layer_idx)] = cer_ic
+
+            loss_interctc = loss_interctc / len(intermediate_outs)
+
+            # calculate whole encoder loss
+            loss_ctc = (
+                               1 - self.interctc_weight
+                       ) * loss_ctc + self.interctc_weight * loss_interctc
+
+        # 2b. Attention decoder branch
+        if self.ctc_weight != 1.0:
+            loss_att, acc_att, cer_att, wer_att = self._calc_att_loss(
+                encoder_out, encoder_out_lens, text, text_lengths
+            )
+
+        # 3. CTC-Att loss definition
+        if self.ctc_weight == 0.0:
+            loss = loss_att
+        elif self.ctc_weight == 1.0:
+            loss = loss_ctc
+        else:
+            loss = self.ctc_weight * loss_ctc + (1 - self.ctc_weight) * loss_att
+
+        # Collect Attn branch stats
+        stats["loss_att"] = loss_att.detach() if loss_att is not None else None
+        stats["acc"] = acc_att
+        stats["cer"] = cer_att
+        stats["wer"] = wer_att
+
+        # Collect total loss stats
+        stats["loss"] = torch.clone(loss.detach())
+
+        # force_gatherable: to-device and to-tensor if scalar for DataParallel
+        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
+        return loss, stats, weight
+
+    def collect_feats(
+            self,
+            speech: torch.Tensor,
+            speech_lengths: torch.Tensor,
+            text: torch.Tensor,
+            text_lengths: torch.Tensor,
+    ) -> Dict[str, torch.Tensor]:
+        if self.extract_feats_in_collect_stats:
+            feats, feats_lengths = self._extract_feats(speech, speech_lengths)
+        else:
+            # Generate dummy stats if extract_feats_in_collect_stats is False
+            logging.warning(
+                "Generating dummy stats for feats and feats_lengths, "
+                "because encoder_conf.extract_feats_in_collect_stats is "
+                f"{self.extract_feats_in_collect_stats}"
+            )
+            feats, feats_lengths = speech, speech_lengths
+        return {"feats": feats, "feats_lengths": feats_lengths}
+
+    def encode(
+            self, speech: torch.Tensor, speech_lengths: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Frontend + Encoder. Note that this method is used by asr_inference.py
+
+        Args:
+            speech: (Batch, Length, ...)
+            speech_lengths: (Batch, )
+        """
+        with autocast(False):
+            # 1. Extract feats
+            feats, feats_lengths = self._extract_feats(speech, speech_lengths)
+
+            # 2. Data augmentation
+            if self.specaug is not None and self.training:
+                feats, feats_lengths = self.specaug(feats, feats_lengths)
+
+            # 3. Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
+            if self.normalize is not None:
+                feats, feats_lengths = self.normalize(feats, feats_lengths)
+
+        # Pre-encoder, e.g. used for raw input data
+        if self.preencoder is not None:
+            feats, feats_lengths = self.preencoder(feats, feats_lengths)
+
+        # 4. Forward encoder
+        # feats: (Batch, Length, Dim)
+        # -> encoder_out: (Batch, Length2, Dim2)
+        if self.encoder.interctc_use_conditioning:
+            encoder_out, encoder_out_lens, _ = self.encoder(
+                feats, feats_lengths, ctc=self.ctc
+            )
+        else:
+            encoder_out, encoder_out_lens, _ = self.encoder(feats, feats_lengths)
+        intermediate_outs = None
+        if isinstance(encoder_out, tuple):
+            intermediate_outs = encoder_out[1]
+            encoder_out = encoder_out[0]
+
+        # Post-encoder, e.g. NLU
+        if self.postencoder is not None:
+            encoder_out, encoder_out_lens = self.postencoder(
+                encoder_out, encoder_out_lens
+            )
+
+        assert encoder_out.size(0) == speech.size(0), (
+            encoder_out.size(),
+            speech.size(0),
+        )
+        assert encoder_out.size(1) <= encoder_out_lens.max(), (
+            encoder_out.size(),
+            encoder_out_lens.max(),
+        )
+
+        if intermediate_outs is not None:
+            return (encoder_out, intermediate_outs), encoder_out_lens
+
+        return encoder_out, encoder_out_lens
+
+    def _extract_feats(
+            self, speech: torch.Tensor, speech_lengths: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        assert speech_lengths.dim() == 1, speech_lengths.shape
+
+        # for data-parallel
+        speech = speech[:, : speech_lengths.max()]
+
+        if self.frontend is not None:
+            # Frontend
+            #  e.g. STFT and Feature extract
+            #       data_loader may send time-domain signal in this case
+            # speech (Batch, NSamples) -> feats: (Batch, NFrames, Dim)
+            feats, feats_lengths = self.frontend(speech, speech_lengths)
+        else:
+            # No frontend and no feature extract
+            feats, feats_lengths = speech, speech_lengths
+        return feats, feats_lengths
+
+    def nll(
+            self,
+            encoder_out: torch.Tensor,
+            encoder_out_lens: torch.Tensor,
+            ys_pad: torch.Tensor,
+            ys_pad_lens: torch.Tensor,
+    ) -> torch.Tensor:
+        """Compute negative log likelihood(nll) from transformer-decoder
+
+        Normally, this function is called in batchify_nll.
+
+        Args:
+            encoder_out: (Batch, Length, Dim)
+            encoder_out_lens: (Batch,)
+            ys_pad: (Batch, Length)
+            ys_pad_lens: (Batch,)
+        """
+        ys_in_pad, ys_out_pad = add_sos_eos(ys_pad, self.sos, self.eos, self.ignore_id)
+        ys_in_lens = ys_pad_lens + 1
+
+        # 1. Forward decoder
+        decoder_out, _ = self.decoder(
+            encoder_out, encoder_out_lens, ys_in_pad, ys_in_lens
+        )  # [batch, seqlen, dim]
+        batch_size = decoder_out.size(0)
+        decoder_num_class = decoder_out.size(2)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            decoder_out.view(-1, decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=self.ignore_id,
+            reduction="none",
+        )
+        nll = nll.view(batch_size, -1)
+        nll = nll.sum(dim=1)
+        assert nll.size(0) == batch_size
+        return nll
+
+    def batchify_nll(
+            self,
+            encoder_out: torch.Tensor,
+            encoder_out_lens: torch.Tensor,
+            ys_pad: torch.Tensor,
+            ys_pad_lens: torch.Tensor,
+            batch_size: int = 100,
+    ):
+        """Compute negative log likelihood(nll) from transformer-decoder
+
+        To avoid OOM, this fuction seperate the input into batches.
+        Then call nll for each batch and combine and return results.
+        Args:
+            encoder_out: (Batch, Length, Dim)
+            encoder_out_lens: (Batch,)
+            ys_pad: (Batch, Length)
+            ys_pad_lens: (Batch,)
+            batch_size: int, samples each batch contain when computing nll,
+                        you may change this to avoid OOM or increase
+                        GPU memory usage
+        """
+        total_num = encoder_out.size(0)
+        if total_num <= batch_size:
+            nll = self.nll(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)
+        else:
+            nll = []
+            start_idx = 0
+            while True:
+                end_idx = min(start_idx + batch_size, total_num)
+                batch_encoder_out = encoder_out[start_idx:end_idx, :, :]
+                batch_encoder_out_lens = encoder_out_lens[start_idx:end_idx]
+                batch_ys_pad = ys_pad[start_idx:end_idx, :]
+                batch_ys_pad_lens = ys_pad_lens[start_idx:end_idx]
+                batch_nll = self.nll(
+                    batch_encoder_out,
+                    batch_encoder_out_lens,
+                    batch_ys_pad,
+                    batch_ys_pad_lens,
+                )
+                nll.append(batch_nll)
+                start_idx = end_idx
+                if start_idx == total_num:
+                    break
+            nll = torch.cat(nll)
+        assert nll.size(0) == total_num
+        return nll
+
+    def _calc_att_loss(
+            self,
+            encoder_out: torch.Tensor,
+            encoder_out_lens: torch.Tensor,
+            ys_pad: torch.Tensor,
+            ys_pad_lens: torch.Tensor,
+    ):
+        ys_in_pad, ys_out_pad = add_sos_eos(ys_pad, self.sos, self.eos, self.ignore_id)
+        ys_in_lens = ys_pad_lens + 1
+
+        # 1. Forward decoder
+        decoder_out, _ = self.decoder(
+            encoder_out, encoder_out_lens, ys_in_pad, ys_in_lens
+        )
+
+        # 2. Compute attention loss
+        loss_att = self.criterion_att(decoder_out, ys_out_pad)
+        acc_att = th_accuracy(
+            decoder_out.view(-1, self.vocab_size),
+            ys_out_pad,
+            ignore_label=self.ignore_id,
+        )
+
+        # Compute cer/wer using attention-decoder
+        if self.training or self.error_calculator is None:
+            cer_att, wer_att = None, None
+        else:
+            ys_hat = decoder_out.argmax(dim=-1)
+            cer_att, wer_att = self.error_calculator(ys_hat.cpu(), ys_pad.cpu())
+
+        return loss_att, acc_att, cer_att, wer_att
+
+    def _calc_ctc_loss(
+            self,
+            encoder_out: torch.Tensor,
+            encoder_out_lens: torch.Tensor,
+            ys_pad: torch.Tensor,
+            ys_pad_lens: torch.Tensor,
+    ):
+        # Calc CTC loss
+        loss_ctc = self.ctc(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)
+
+        # Calc CER using CTC
+        cer_ctc = None
+        if not self.training and self.error_calculator is not None:
+            ys_hat = self.ctc.argmax(encoder_out).data
+            cer_ctc = self.error_calculator(ys_hat.cpu(), ys_pad.cpu(), is_ctc=True)
+        return loss_ctc, cer_ctc

--
Gitblit v1.9.1