diff --git a/funasr/models/paraformer/decoder.py b/funasr/models/paraformer/decoder.py index 68018a039..ad321e4f4 100644 --- a/funasr/models/paraformer/decoder.py +++ b/funasr/models/paraformer/decoder.py @@ -116,6 +116,22 @@ class DecoderLayerSANM(torch.nn.Module): # x = residual + self.dropout(self.src_attn(x, memory, memory_mask)) return x, tgt_mask, memory, memory_mask, cache + + def get_attn_mat(self, tgt, tgt_mask, memory, memory_mask=None, cache=None): + residual = tgt + tgt = self.norm1(tgt) + tgt = self.feed_forward(tgt) + + x = tgt + if self.self_attn is not None: + tgt = self.norm2(tgt) + x, cache = self.self_attn(tgt, tgt_mask, cache=cache) + x = residual + x + + residual = x + x = self.norm3(x) + x_src_attn, attn_mat = self.src_attn(x, memory, memory_mask, ret_attn=True) + return attn_mat def forward_one_step(self, tgt, tgt_mask, memory, memory_mask=None, cache=None): """Compute decoded features. @@ -396,6 +412,46 @@ class ParaformerSANMDecoder(BaseTransformerDecoder): ys.unsqueeze(0), ys_mask, x.unsqueeze(0), cache=state ) return logp.squeeze(0), state + + def forward_asf2( + self, + hs_pad: torch.Tensor, + hlens: torch.Tensor, + ys_in_pad: torch.Tensor, + ys_in_lens: torch.Tensor, + ): + + tgt = ys_in_pad + tgt_mask = myutils.sequence_mask(ys_in_lens, device=tgt.device)[:, :, None] + + memory = hs_pad + memory_mask = myutils.sequence_mask(hlens, device=memory.device)[:, None, :] + + tgt, tgt_mask, memory, memory_mask, _ = self.decoders[0](tgt, tgt_mask, memory, memory_mask) + attn_mat = self.model.decoders[1].get_attn_mat(tgt, tgt_mask, memory, memory_mask) + return attn_mat + + def forward_asf6( + self, + hs_pad: torch.Tensor, + hlens: torch.Tensor, + ys_in_pad: torch.Tensor, + ys_in_lens: torch.Tensor, + ): + + tgt = ys_in_pad + tgt_mask = myutils.sequence_mask(ys_in_lens, device=tgt.device)[:, :, None] + + memory = hs_pad + memory_mask = myutils.sequence_mask(hlens, device=memory.device)[:, None, :] + + tgt, tgt_mask, memory, memory_mask, _ = self.decoders[0](tgt, tgt_mask, memory, memory_mask) + tgt, tgt_mask, memory, memory_mask, _ = self.decoders[1](tgt, tgt_mask, memory, memory_mask) + tgt, tgt_mask, memory, memory_mask, _ = self.decoders[2](tgt, tgt_mask, memory, memory_mask) + tgt, tgt_mask, memory, memory_mask, _ = self.decoders[3](tgt, tgt_mask, memory, memory_mask) + tgt, tgt_mask, memory, memory_mask, _ = self.decoders[4](tgt, tgt_mask, memory, memory_mask) + attn_mat = self.decoders[5].get_attn_mat(tgt, tgt_mask, memory, memory_mask) + return attn_mat def forward_chunk( self, diff --git a/funasr/models/seaco_paraformer/model.py b/funasr/models/seaco_paraformer/model.py index caf2b15c7..cfdd26a79 100644 --- a/funasr/models/seaco_paraformer/model.py +++ b/funasr/models/seaco_paraformer/model.py @@ -19,11 +19,9 @@ from distutils.version import LooseVersion from funasr.register import tables from funasr.utils import postprocess_utils -from funasr.metrics.compute_acc import th_accuracy from funasr.models.paraformer.model import Paraformer from funasr.utils.datadir_writer import DatadirWriter from funasr.models.paraformer.search import Hypothesis -from funasr.models.paraformer.cif_predictor import mae_loss 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 @@ -76,7 +74,7 @@ class SeacoParaformer(BiCifParaformer, Paraformer): self.lstm_proj = torch.nn.Linear(self.inner_dim*2, self.inner_dim) else: self.lstm_proj = None - self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim) + # self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim) elif self.bias_encoder_type == 'mean': self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim) else: @@ -225,12 +223,8 @@ class SeacoParaformer(BiCifParaformer, Paraformer): # 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 = self.seaco_decoder.forward_asf6(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens) + hotword_scores = hotword_scores[0].sum(0).sum(0) # 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() add_filter = dec_filter @@ -241,9 +235,6 @@ class SeacoParaformer(BiCifParaformer, Paraformer): 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) - 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) @@ -274,6 +265,8 @@ class SeacoParaformer(BiCifParaformer, Paraformer): hotword_lengths): if self.bias_encoder_type != 'lstm': logging.error("Unsupported bias encoder type") + + ''' hw_embed = self.decoder.embed(hotword_pad) hw_embed, (_, _) = self.bias_encoder(hw_embed) if self.lstm_proj is not None: @@ -281,26 +274,20 @@ class SeacoParaformer(BiCifParaformer, Paraformer): _ind = np.arange(0, hw_embed.shape[0]).tolist() selected = hw_embed[_ind, [i-1 for i in hotword_lengths.detach().cpu().tolist()]] return selected + ''' - ''' - 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) - pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index, pre_token_length2 = self.predictor(encoder_out, - None, - encoder_out_mask, - ignore_id=self.ignore_id) - return pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index - - - def calc_predictor_timestamp(self, encoder_out, encoder_out_lens, token_num): - encoder_out_mask = (~make_pad_mask(encoder_out_lens, maxlen=encoder_out.size(1))[:, None, :]).to( - encoder_out.device) - ds_alphas, ds_cif_peak, us_alphas, us_peaks = self.predictor.get_upsample_timestamp(encoder_out, - encoder_out_mask, - token_num) - return ds_alphas, ds_cif_peak, us_alphas, us_peaks - ''' + # hw_embed = self.sac_embedding(hotword_pad) + hw_embed = self.decoder.embed(hotword_pad) + hw_embed = torch.nn.utils.rnn.pack_padded_sequence(hw_embed, hotword_lengths.cpu().type(torch.int64), batch_first=True, enforce_sorted=False) + packed_rnn_output, _ = self.bias_encoder(hw_embed) + rnn_output = torch.nn.utils.rnn.pad_packed_sequence(packed_rnn_output, batch_first=True)[0] + if self.lstm_proj is not None: + hw_hidden = self.lstm_proj(rnn_output) + else: + hw_hidden = rnn_output + _ind = np.arange(0, hw_hidden.shape[0]).tolist() + selected = hw_hidden[_ind, [i-1 for i in hotword_lengths.detach().cpu().tolist()]] + return selected def inference(self, data_in,