Merge pull request #364 from alibaba-damo-academy/dev_sx

Lizerui9926

2023-04-17 39efd2204c84b086dceac3dbb257425fc640a3ff

 funasr/models/e2e_asr_paraformer.py

@@ -44,7 +44,7 @@

class Paraformer(AbsESPnetModel):
    """
    Author: Speech Lab, Alibaba Group, China
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
    https://arxiv.org/abs/2206.08317
    """
@@ -370,19 +370,10 @@
                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
        return encoder_out, torch.tensor([encoder_out.size(1)])

    def calc_predictor(self, encoder_out, encoder_out_lens):

@@ -621,7 +612,7 @@

class ParaformerBert(Paraformer):
    """
    Author: Speech Lab, Alibaba Group, China
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Paraformer2: advanced paraformer with LFMMI and bert for non-autoregressive end-to-end speech recognition
    """

@@ -986,6 +977,59 @@
        loss_pre2 = self.criterion_pre(ys_pad_lens.type_as(pre_token_length2), pre_token_length2)

        return loss_pre2

    def _calc_att_loss(
            self,
            encoder_out: torch.Tensor,
            encoder_out_lens: torch.Tensor,
            ys_pad: torch.Tensor,
            ys_pad_lens: torch.Tensor,
    ):
        encoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to(
            encoder_out.device)
        if self.predictor_bias == 1:
            _, ys_pad = add_sos_eos(ys_pad, self.sos, self.eos, self.ignore_id)
            ys_pad_lens = ys_pad_lens + self.predictor_bias
        pre_acoustic_embeds, pre_token_length, _, pre_peak_index, _ = self.predictor(encoder_out, ys_pad, encoder_out_mask,
                                                                                  ignore_id=self.ignore_id)

        # 0. sampler
        decoder_out_1st = None
        if self.sampling_ratio > 0.0:
            if self.step_cur < 2:
                logging.info("enable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))
            sematic_embeds, decoder_out_1st = self.sampler(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens,
                                                           pre_acoustic_embeds)
        else:
            if self.step_cur < 2:
                logging.info("disable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))
            sematic_embeds = pre_acoustic_embeds

        # 1. Forward decoder
        decoder_outs = self.decoder(
            encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens
        )
        decoder_out, _ = decoder_outs[0], decoder_outs[1]

        if decoder_out_1st is None:
            decoder_out_1st = decoder_out
        # 2. Compute attention loss
        loss_att = self.criterion_att(decoder_out, ys_pad)
        acc_att = th_accuracy(
            decoder_out_1st.view(-1, self.vocab_size),
            ys_pad,
            ignore_label=self.ignore_id,
        )
        loss_pre = self.criterion_pre(ys_pad_lens.type_as(pre_token_length), pre_token_length)

        # 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_1st.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, loss_pre
    
    def calc_predictor(self, encoder_out, encoder_out_lens):

@@ -1034,16 +1078,76 @@

        # 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
        loss_pre = 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, loss_pre = self._calc_att_loss(
                encoder_out, encoder_out_lens, text, text_lengths
            )

        loss_pre2 = self._calc_pre2_loss(
            encoder_out, encoder_out_lens, text, text_lengths
        )

        loss = loss_pre2
        # 3. CTC-Att loss definition
        if self.ctc_weight == 0.0:
            loss = loss_att + loss_pre * self.predictor_weight + loss_pre2 * self.predictor_weight * 0.5
        elif self.ctc_weight == 1.0:
            loss = loss_ctc
        else:
            loss = self.ctc_weight * loss_ctc + (1 - self.ctc_weight) * loss_att + loss_pre * self.predictor_weight + loss_pre2 * self.predictor_weight * 0.5

        # 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
        stats["loss_pre"] = loss_pre.detach().cpu() if loss_pre is not None else None
        stats["loss_pre2"] = loss_pre2.detach().cpu()

        stats["loss"] = torch.clone(loss.detach())

        # force_gatherable: to-device and to-tensor if scalar for DataParallel
@@ -1094,6 +1198,7 @@
            inner_dim: int = 256,
            bias_encoder_type: str = 'lstm',
            label_bracket: bool = False,
            use_decoder_embedding: bool = False,
    ):
        assert check_argument_types()
        assert 0.0 <= ctc_weight <= 1.0, ctc_weight
@@ -1147,6 +1252,7 @@
            self.hotword_buffer = None
            self.length_record = []
            self.current_buffer_length = 0
        self.use_decoder_embedding = use_decoder_embedding

    def forward(
            self,
@@ -1288,7 +1394,10 @@
                    hw_list.append(hw_tokens)
        # padding
        hw_list_pad = pad_list(hw_list, 0)
        hw_embed = self.decoder.embed(hw_list_pad)
        if self.use_decoder_embedding:
            hw_embed = self.decoder.embed(hw_list_pad)
        else:
            hw_embed = self.bias_embed(hw_list_pad)
        hw_embed, (_, _) = self.bias_encoder(hw_embed)
        _ind = np.arange(0, len(hw_list)).tolist()
        # update self.hotword_buffer, throw a part if oversize
@@ -1404,13 +1513,19 @@
            # default hotword list
            hw_list = [torch.Tensor([self.sos]).long().to(encoder_out.device)]  # empty hotword list
            hw_list_pad = pad_list(hw_list, 0)
            hw_embed = self.bias_embed(hw_list_pad)
            if self.use_decoder_embedding:
                hw_embed = self.decoder.embed(hw_list_pad)
            else:
                hw_embed = self.bias_embed(hw_list_pad)
            _, (h_n, _) = self.bias_encoder(hw_embed)
            contextual_info = h_n.squeeze(0).repeat(encoder_out.shape[0], 1, 1)
        else:
            hw_lengths = [len(i) for i in hw_list]
            hw_list_pad = pad_list([torch.Tensor(i).long() for i in hw_list], 0).to(encoder_out.device)
            hw_embed = self.bias_embed(hw_list_pad)
            if self.use_decoder_embedding:
                hw_embed = self.decoder.embed(hw_list_pad)
            else:
                hw_embed = self.bias_embed(hw_list_pad)
            hw_embed = torch.nn.utils.rnn.pack_padded_sequence(hw_embed, hw_lengths, batch_first=True,
                                                               enforce_sorted=False)
            _, (h_n, _) = self.bias_encoder(hw_embed)