From 54931dd4e1a099d7d6f144c4e12e5453deb3aa26 Mon Sep 17 00:00:00 2001
From: 雾聪 <wucong.lyb@alibaba-inc.com>
Date: 星期三, 28 六月 2023 10:41:57 +0800
Subject: [PATCH] Merge branch 'main' of https://github.com/alibaba-damo-academy/FunASR into main
---
funasr/models/e2e_asr_contextual_paraformer.py | 35 ++++++-----------------------------
1 files changed, 6 insertions(+), 29 deletions(-)
diff --git a/funasr/models/e2e_asr_contextual_paraformer.py b/funasr/models/e2e_asr_contextual_paraformer.py
index 493b345..cfb5008 100644
--- a/funasr/models/e2e_asr_contextual_paraformer.py
+++ b/funasr/models/e2e_asr_contextual_paraformer.py
@@ -43,9 +43,7 @@
frontend: Optional[AbsFrontend],
specaug: Optional[AbsSpecAug],
normalize: Optional[AbsNormalize],
- preencoder: Optional[AbsPreEncoder],
encoder: AbsEncoder,
- postencoder: Optional[AbsPostEncoder],
decoder: AbsDecoder,
ctc: CTC,
ctc_weight: float = 0.5,
@@ -72,6 +70,8 @@
crit_attn_weight: float = 0.0,
crit_attn_smooth: float = 0.0,
bias_encoder_dropout_rate: float = 0.0,
+ preencoder: Optional[AbsPreEncoder] = None,
+ postencoder: Optional[AbsPostEncoder] = None,
):
assert check_argument_types()
assert 0.0 <= ctc_weight <= 1.0, ctc_weight
@@ -280,8 +280,8 @@
decoder_outs = self.decoder(
encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, contextual_info=contextual_info
)
- decoder_out, _, attn = decoder_outs[0], decoder_outs[1], decoder_outs[2]
-
+ decoder_out, _ = decoder_outs[0], decoder_outs[1]
+ '''
if self.crit_attn_weight > 0 and attn.shape[-1] > 1:
ideal_attn = ideal_attn + self.crit_attn_smooth / (self.crit_attn_smooth + 1.0)
attn_non_blank = attn[:,:,:,:-1]
@@ -289,6 +289,8 @@
loss_ideal = self.attn_loss(attn_non_blank.max(1)[0], ideal_attn_non_blank.to(attn.device))
else:
loss_ideal = None
+ '''
+ loss_ideal = None
if decoder_out_1st is None:
decoder_out_1st = decoder_out
@@ -360,11 +362,6 @@
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)
- # hw_embed, _ = torch.nn.utils.rnn.pad_packed_sequence(hw_embed, batch_first=True)
- if h_n.shape[1] > 2000: # large hotword list
- _h_n = self.pick_hwlist_group(h_n.squeeze(0), encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens)
- if _h_n is not None:
- h_n = _h_n
hw_embed = h_n.repeat(encoder_out.shape[0], 1, 1)
decoder_outs = self.decoder(
@@ -373,23 +370,3 @@
decoder_out = decoder_outs[0]
decoder_out = torch.log_softmax(decoder_out, dim=-1)
return decoder_out, ys_pad_lens
-
- def pick_hwlist_group(self, hw_embed, encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens):
- max_attn_score = 0.0
- # max_attn_index = 0
- argmax_g = None
- non_blank = hw_embed[-1]
- hw_embed_groups = hw_embed[:-1].split(2000)
- for i, g in enumerate(hw_embed_groups):
- g = torch.cat([g, non_blank.unsqueeze(0)], dim=0)
- _ = self.decoder(
- encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, contextual_info=g.unsqueeze(0)
- )
- attn = self.decoder.bias_decoder.src_attn.attn[0]
- _max_attn_score = attn.max(0)[0][:,:-1].max()
- if _max_attn_score > max_attn_score:
- max_attn_score = _max_attn_score
- # max_attn_index = i
- argmax_g = g
- # import pdb; pdb.set_trace()
- return argmax_g
\ No newline at end of file
--
Gitblit v1.9.1