Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

update asf code

shixian.shi5b38115da4f9
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 0287f56ed..cfdd26a79 100644
--- a/funasr/models/seaco_paraformer/model.py
+++ b/funasr/models/seaco_paraformer/model.py
@@ -223,12 +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
@@ -239,9 +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)