From c2dee5e3c29eba79e591d9e9caebaef15ea4e56b Mon Sep 17 00:00:00 2001
From: hnluo <haoneng.lhn@alibaba-inc.com>
Date: 星期四, 29 六月 2023 11:09:28 +0800
Subject: [PATCH] Merge pull request #687 from alibaba-damo-academy/dev_lhn
---
funasr/models/decoder/sanm_decoder.py | 24 ++++++++++++++----------
1 files changed, 14 insertions(+), 10 deletions(-)
diff --git a/funasr/models/decoder/sanm_decoder.py b/funasr/models/decoder/sanm_decoder.py
index 3bfcffc..ed920bf 100644
--- a/funasr/models/decoder/sanm_decoder.py
+++ b/funasr/models/decoder/sanm_decoder.py
@@ -104,7 +104,6 @@
x = residual + self.dropout(self.src_attn(x, memory, memory_mask))
-
return x, tgt_mask, memory, memory_mask, cache
def forward_chunk(self, tgt, tgt_mask, memory, memory_mask=None, cache=None):
@@ -152,7 +151,7 @@
class FsmnDecoderSCAMAOpt(BaseTransformerDecoder):
"""
- author: Speech Lab, Alibaba Group, China
+ Author: Speech Lab of DAMO Academy, Alibaba Group
SCAMA: Streaming chunk-aware multihead attention for online end-to-end speech recognition
https://arxiv.org/abs/2006.01713
@@ -400,7 +399,7 @@
for i in range(self.att_layer_num):
decoder = self.decoders[i]
c = cache[i]
- x, tgt_mask, memory, memory_mask, c_ret = decoder(
+ x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
x, tgt_mask, memory, memory_mask, cache=c
)
new_cache.append(c_ret)
@@ -410,13 +409,13 @@
j = i + self.att_layer_num
decoder = self.decoders2[i]
c = cache[j]
- x, tgt_mask, memory, memory_mask, c_ret = decoder(
+ x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
x, tgt_mask, memory, memory_mask, cache=c
)
new_cache.append(c_ret)
for decoder in self.decoders3:
- x, tgt_mask, memory, memory_mask, _ = decoder(
+ x, tgt_mask, memory, memory_mask, _ = decoder.forward_chunk(
x, tgt_mask, memory, None, cache=None
)
@@ -813,7 +812,7 @@
class ParaformerSANMDecoder(BaseTransformerDecoder):
"""
- author: Speech Lab, Alibaba Group, China
+ Author: Speech Lab of DAMO Academy, Alibaba Group
Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
https://arxiv.org/abs/2006.01713
"""
@@ -936,6 +935,7 @@
hlens: torch.Tensor,
ys_in_pad: torch.Tensor,
ys_in_lens: torch.Tensor,
+ chunk_mask: torch.Tensor = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Forward decoder.
@@ -956,9 +956,13 @@
"""
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, :]
+ if chunk_mask is not None:
+ memory_mask = memory_mask * chunk_mask
+ if tgt_mask.size(1) != memory_mask.size(1):
+ memory_mask = torch.cat((memory_mask, memory_mask[:, -2:-1, :]), dim=1)
x = tgt
x, tgt_mask, memory, memory_mask, _ = self.decoders(
@@ -1077,7 +1081,7 @@
for i in range(self.att_layer_num):
decoder = self.decoders[i]
c = cache[i]
- x, tgt_mask, memory, memory_mask, c_ret = decoder(
+ x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
x, tgt_mask, memory, None, cache=c
)
new_cache.append(c_ret)
@@ -1087,14 +1091,14 @@
j = i + self.att_layer_num
decoder = self.decoders2[i]
c = cache[j]
- x, tgt_mask, memory, memory_mask, c_ret = decoder(
+ x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_chunk(
x, tgt_mask, memory, None, cache=c
)
new_cache.append(c_ret)
for decoder in self.decoders3:
- x, tgt_mask, memory, memory_mask, _ = decoder(
+ x, tgt_mask, memory, memory_mask, _ = decoder.forward_chunk(
x, tgt_mask, memory, None, cache=None
)
--
Gitblit v1.9.1