python/FunASR-XL.git

			@@ -32,7 +32,7 @@
			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


			import pdb
			if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):
			from torch.cuda.amp import autocast
			else:
			@@ -66,7 +66,6 @@

			# bias encoder
			if self.bias_encoder_type == 'lstm':
			logging.warning("enable bias encoder sampling and contextual training")
			self.bias_encoder = torch.nn.LSTM(self.inner_dim,
			self.inner_dim,
			2,
			@@ -79,7 +78,6 @@
			self.lstm_proj = None
			self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim)
			elif self.bias_encoder_type == 'mean':
			logging.warning("enable bias encoder sampling and contextual training")
			self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim)
			else:
			logging.error("Unsupport bias encoder type: {}".format(self.bias_encoder_type))
			@@ -132,7 +130,7 @@
			hotword_pad = kwargs.get("hotword_pad")
			hotword_lengths = kwargs.get("hotword_lengths")
			dha_pad = kwargs.get("dha_pad")


			batch_size = speech.shape[0]
			self.step_cur += 1
			# for data-parallel
			@@ -214,24 +212,29 @@
			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:
			for dec in self.seaco_decoder.decoders:
			dec.reserve_attn = True

			# cif_attended, _ = self.decoder2(contextual_info, _contextual_length, sematic_embeds, ys_pad_lens)
			dec_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
			# cif_filter = torch.topk(self.decoder2.decoders[-1].attn_mat[0][0].sum(0).sum(0)[:-1], min(nfilter, num_hot_word-1))[1].tolist()

			hotword_scores = self.seaco_decoder.decoders[-1].attn_mat[0][0].sum(0).sum(0)[:-1]
			# hotword_scores /= torch.sqrt(torch.tensor(hw_lengths)[:-1].float()).to(hotword_scores.device)
			dec_filter = torch.topk(hotword_scores, min(nfilter, num_hot_word-1))[1].tolist()
			@@ -246,12 +249,11 @@
			for dec in self.seaco_decoder.decoders:
			dec.attn_mat = []
			dec.reserve_attn = False

			# SeACo Core
			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):
			@@ -265,8 +267,8 @@
			# 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
			@@ -320,7 +322,6 @@
			logging.info("enable beam_search")
			self.init_beam_search(**kwargs)
			self.nbest = kwargs.get("nbest", 1)

			meta_data = {}

			# extract fbank feats
			@@ -337,7 +338,7 @@

			speech = speech.to(device=kwargs["device"])
			speech_lengths = speech_lengths.to(device=kwargs["device"])


			# hotword
			self.hotword_list = self.generate_hotwords_list(kwargs.get("hotword", None), tokenizer=tokenizer, frontend=frontend)

			@@ -346,6 +347,7 @@
			if isinstance(encoder_out, tuple):
			encoder_out = encoder_out[0]


			# predictor
			predictor_outs = self.calc_predictor(encoder_out, encoder_out_lens)
			pre_acoustic_embeds, pre_token_length, _, _ = predictor_outs[0], predictor_outs[1], \
			@@ -354,15 +356,14 @@
			if torch.max(pre_token_length) < 1:
			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, _ = decoder_outs[0], decoder_outs[1]
			_, _, us_alphas, us_peaks = self.calc_predictor_timestamp(encoder_out, encoder_out_lens,
			pre_token_length)

			results = []
			b, n, d = decoder_out.size()
			for i in range(b):
			@@ -387,9 +388,11 @@
			nbest_hyps = [Hypothesis(yseq=yseq, score=score)]
			for nbest_idx, hyp in enumerate(nbest_hyps):
			ibest_writer = None
			if ibest_writer is None and kwargs.get("output_dir") is not None:
			writer = DatadirWriter(kwargs.get("output_dir"))
			ibest_writer = writer[f"{nbest_idx + 1}best_recog"]
			if kwargs.get("output_dir") is not None:
			if not hasattr(self, "writer"):
			self.writer = DatadirWriter(kwargs.get("output_dir"))
			ibest_writer = self.writer[f"{nbest_idx + 1}best_recog"]

			# remove sos/eos and get results
			last_pos = -1
			if isinstance(hyp.yseq, list):
			@@ -415,12 +418,11 @@
			token, timestamp)

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

			if ibest_writer is not None:
			ibest_writer["token"][key[i]] = " ".join(token)
			# ibest_writer["text"][key[i]] = text
			ibest_writer["timestamp"][key[i]] = time_stamp_postprocessed
			ibest_writer["text"][key[i]] = text_postprocessed
			else: