python/FunASR-XL.git

			@@ -25,8 +25,8 @@
			from funasr.train_utils.device_funcs import force_gatherable
			from funasr.models.bicif_paraformer.model import BiCifParaformer
			from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
			from funasr.utils.timestamp_tools import ts_prediction_lfr6_standard
			from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
			from funasr.utils.timestamp_tools import ts_prediction_lfr6_standard
			from funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_list
			from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank

			@@ -97,8 +97,10 @@
			smoothing=seaco_lsm_weight,
			normalize_length=seaco_length_normalized_loss,
			)
			self.train_decoder = kwargs.get("train_decoder", False)
			self.train_decoder = kwargs.get("train_decoder", True)
			self.seaco_weight = kwargs.get("seaco_weight", 0.01)
			self.NO_BIAS = kwargs.get("NO_BIAS", 8377)
			self.predictor_name = kwargs.get("predictor")

			def forward(
			self,
			@@ -116,7 +118,10 @@
			text: (Batch, Length)
			text_lengths: (Batch,)
			"""
			assert text_lengths.dim() == 1, text_lengths.shape
			if len(text_lengths.size()) > 1:
			text_lengths = text_lengths[:, 0]
			if len(speech_lengths.size()) > 1:
			speech_lengths = speech_lengths[:, 0]
			# Check that batch_size is unified
			assert (
			speech.shape[0]
			@@ -127,10 +132,11 @@

			hotword_pad = kwargs.get("hotword_pad")
			hotword_lengths = kwargs.get("hotword_lengths")
			dha_pad = kwargs.get("dha_pad")

			seaco_label_pad = kwargs.get("seaco_label_pad")
			if len(hotword_lengths.size()) > 1:
			hotword_lengths = hotword_lengths[:, 0]

			batch_size = speech.shape[0]
			self.step_cur += 1
			# for data-parallel
			text = text[:, : text_lengths.max()]
			speech = speech[:, :speech_lengths.max()]
			@@ -148,28 +154,35 @@
			ys_lengths,
			hotword_pad,
			hotword_lengths,
			dha_pad,
			seaco_label_pad,
			)
			if self.train_decoder:
			loss_att, acc_att = self._calc_att_loss(
			loss_att, acc_att, _, _, _ = self._calc_att_loss(
			encoder_out, encoder_out_lens, text, text_lengths
			)
			loss = loss_seaco + loss_att
			loss = loss_seaco + loss_att * self.seaco_weight
			stats["loss_att"] = torch.clone(loss_att.detach())
			stats["acc_att"] = acc_att
			else:
			loss = loss_seaco

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

			# force_gatherable: to-device and to-tensor if scalar for DataParallel
			if self.length_normalized_loss:
			batch_size = (text_lengths + self.predictor_bias).sum().type_as(batch_size)
			batch_size = (text_lengths + self.predictor_bias).sum()
			loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
			return loss, stats, weight

			def _merge(self, cif_attended, dec_attended):
			return cif_attended + dec_attended

			def calc_predictor(self, encoder_out, encoder_out_lens):
			encoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to(
			encoder_out.device)
			predictor_outs = self.predictor(encoder_out, None, encoder_out_mask, ignore_id=self.ignore_id)
			return predictor_outs[:4]

			def _calc_seaco_loss(
			self,
			@@ -179,13 +192,12 @@
			ys_lengths: torch.Tensor,
			hotword_pad: torch.Tensor,
			hotword_lengths: torch.Tensor,
			dha_pad: torch.Tensor,
			seaco_label_pad: torch.Tensor,
			):
			# predictor forward
			encoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to(
			encoder_out.device)
			pre_acoustic_embeds, _, _, _ = self.predictor(encoder_out, ys_pad, encoder_out_mask,
			ignore_id=self.ignore_id)
			pre_acoustic_embeds = self.predictor(encoder_out, ys_pad, encoder_out_mask, ignore_id=self.ignore_id)[0]
			# decoder forward
			decoder_out, _ = self.decoder(encoder_out, encoder_out_lens, pre_acoustic_embeds, ys_lengths, return_hidden=True)
			selected = self._hotword_representation(hotword_pad,
			@@ -198,7 +210,7 @@
			dec_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, decoder_out, ys_lengths)
			merged = self._merge(cif_attended, dec_attended)
			dha_output = self.hotword_output_layer(merged[:, :-1]) # remove the last token in loss calculation
			loss_att = self.criterion_seaco(dha_output, dha_pad)
			loss_att = self.criterion_seaco(dha_output, seaco_label_pad)
			return loss_att

			def _seaco_decode_with_ASF(self,
			@@ -210,17 +222,20 @@
			nfilter=50,
			seaco_weight=1.0):
			# decoder forward

			decoder_out, decoder_hidden, _ = self.decoder(encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, return_hidden=True, return_both=True)

			decoder_pred = torch.log_softmax(decoder_out, dim=-1)
			if hw_list is not None:
			hw_lengths = [len(i) for i in hw_list]
			hw_list_ = [torch.Tensor(i).long() for i in hw_list]
			hw_list_pad = pad_list(hw_list_, 0).to(encoder_out.device)
			selected = self._hotword_representation(hw_list_pad, torch.Tensor(hw_lengths).int().to(encoder_out.device))

			contextual_info = selected.squeeze(0).repeat(encoder_out.shape[0], 1, 1).to(encoder_out.device)
			num_hot_word = contextual_info.shape[1]
			_contextual_length = torch.Tensor([num_hot_word]).int().repeat(encoder_out.shape[0]).to(encoder_out.device)


			# ASF Core
			if nfilter > 0 and nfilter < num_hot_word:
			hotword_scores = self.seaco_decoder.forward_asf6(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
			@@ -240,13 +255,13 @@
			cif_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, sematic_embeds, ys_pad_lens)
			dec_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
			merged = self._merge(cif_attended, dec_attended)


			dha_output = self.hotword_output_layer(merged) # remove the last token in loss calculation
			dha_pred = torch.log_softmax(dha_output, dim=-1)
			def _merge_res(dec_output, dha_output):
			lmbd = torch.Tensor([seaco_weight] * dha_output.shape[0])
			dha_ids = dha_output.max(-1)[-1]# [0]
			dha_mask = (dha_ids == 8377).int().unsqueeze(-1)
			dha_mask = (dha_ids == self.NO_BIAS).int().unsqueeze(-1)
			a = (1 - lmbd) / lmbd
			b = 1 / lmbd
			a, b = a.to(dec_output.device), b.to(dec_output.device)
			@@ -254,8 +269,8 @@
			# logits = dec_output * dha_mask + dha_output[:,:,:-1] * (1-dha_mask)
			logits = dec_output * dha_mask + dha_output[:,:,:] * (1-dha_mask)
			return logits

			merged_pred = _merge_res(decoder_pred, dha_pred)
			# import pdb; pdb.set_trace()
			return merged_pred
			else:
			return decoder_pred
			@@ -305,7 +320,6 @@
			logging.info("enable beam_search")
			self.init_beam_search(**kwargs)
			self.nbest = kwargs.get("nbest", 1)

			meta_data = {}

			# extract fbank feats
			@@ -333,21 +347,26 @@

			# predictor
			predictor_outs = self.calc_predictor(encoder_out, encoder_out_lens)
			pre_acoustic_embeds, pre_token_length, _, _ = predictor_outs[0], predictor_outs[1], \
			predictor_outs[2], predictor_outs[3]
			pre_acoustic_embeds, pre_token_length = predictor_outs[0], predictor_outs[1]
			pre_token_length = pre_token_length.round().long()
			if torch.max(pre_token_length) < 1:
			return []
			return ([],)

			decoder_out = self._seaco_decode_with_ASF(encoder_out,
			encoder_out_lens,
			pre_acoustic_embeds,
			pre_token_length,
			hw_list=self.hotword_list
			)

			decoder_out = self._seaco_decode_with_ASF(encoder_out, encoder_out_lens,
			pre_acoustic_embeds,
			pre_token_length,
			hw_list=self.hotword_list)
			# decoder_out, _ = decoder_outs[0], decoder_outs[1]
			_, _, us_alphas, us_peaks = self.calc_predictor_timestamp(encoder_out, encoder_out_lens,
			pre_token_length)

			if self.predictor_name == "CifPredictorV3":
			_, _, us_alphas, us_peaks = self.calc_predictor_timestamp(encoder_out,
			encoder_out_lens,
			pre_token_length)
			else:
			us_alphas = None

			results = []
			b, n, d = decoder_out.size()
			for i in range(b):
			@@ -392,29 +411,30 @@
			# Change integer-ids to tokens
			token = tokenizer.ids2tokens(token_int)
			text = tokenizer.tokens2text(token)

			_, timestamp = ts_prediction_lfr6_standard(us_alphas[i][:encoder_out_lens[i] * 3],
			us_peaks[i][:encoder_out_lens[i] * 3],
			copy.copy(token),
			vad_offset=kwargs.get("begin_time", 0))

			text_postprocessed, time_stamp_postprocessed, word_lists = postprocess_utils.sentence_postprocess(
			token, timestamp)

			result_i = {"key": key[i], "text": text_postprocessed,
			"timestamp": time_stamp_postprocessed
			}

			if ibest_writer is not None:
			ibest_writer["token"][key[i]] = " ".join(token)
			ibest_writer["timestamp"][key[i]] = time_stamp_postprocessed
			ibest_writer["text"][key[i]] = text_postprocessed
			if us_alphas is not None:
			_, timestamp = ts_prediction_lfr6_standard(us_alphas[i][:encoder_out_lens[i] * 3],
			us_peaks[i][:encoder_out_lens[i] * 3],
			copy.copy(token),
			vad_offset=kwargs.get("begin_time", 0))
			text_postprocessed, time_stamp_postprocessed, _ = \
			postprocess_utils.sentence_postprocess(token, timestamp)
			result_i = {"key": key[i], "text": text_postprocessed,
			"timestamp": time_stamp_postprocessed}
			if ibest_writer is not None:
			ibest_writer["token"][key[i]] = " ".join(token)
			ibest_writer["timestamp"][key[i]] = time_stamp_postprocessed
			ibest_writer["text"][key[i]] = text_postprocessed
			else:
			text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
			result_i = {"key": key[i], "text": text_postprocessed}
			if ibest_writer is not None:
			ibest_writer["token"][key[i]] = " ".join(token)
			ibest_writer["text"][key[i]] = text_postprocessed
			else:
			result_i = {"key": key[i], "token_int": token_int}
			results.append(result_i)

			return results, meta_data


			def generate_hotwords_list(self, hotword_list_or_file, tokenizer=None, frontend=None):
			def load_seg_dict(seg_dict_file):
			@@ -517,3 +537,13 @@
			hotword_list = None
			return hotword_list

			def export(
			self,
			**kwargs,
			):
			if 'max_seq_len' not in kwargs:
			kwargs['max_seq_len'] = 512
			from .export_meta import export_rebuild_model
			models = export_rebuild_model(model=self, **kwargs)
			return models