From 219c2482ab755fbd4e49dfbdee91bf1a8a4ec49a Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 19 五月 2023 11:33:27 +0800
Subject: [PATCH] websocket 2pass bugfix
---
funasr/models/encoder/conformer_encoder.py | 15 +++++++++------
1 files changed, 9 insertions(+), 6 deletions(-)
diff --git a/funasr/models/encoder/conformer_encoder.py b/funasr/models/encoder/conformer_encoder.py
index b7b552c..5f20dee 100644
--- a/funasr/models/encoder/conformer_encoder.py
+++ b/funasr/models/encoder/conformer_encoder.py
@@ -15,13 +15,13 @@
from typeguard import check_argument_types
from funasr.models.ctc import CTC
-from funasr.models.encoder.abs_encoder import AbsEncoder
from funasr.modules.attention import (
MultiHeadedAttention, # noqa: H301
RelPositionMultiHeadedAttention, # noqa: H301
RelPositionMultiHeadedAttentionChunk,
LegacyRelPositionMultiHeadedAttention, # noqa: H301
)
+from funasr.models.encoder.abs_encoder import AbsEncoder
from funasr.modules.embedding import (
PositionalEncoding, # noqa: H301
ScaledPositionalEncoding, # noqa: H301
@@ -307,7 +307,7 @@
feed_forward: torch.nn.Module,
feed_forward_macaron: torch.nn.Module,
conv_mod: torch.nn.Module,
- norm_class: torch.nn.Module = torch.nn.LayerNorm,
+ norm_class: torch.nn.Module = LayerNorm,
norm_args: Dict = {},
dropout_rate: float = 0.0,
) -> None:
@@ -894,7 +894,7 @@
return x, cache
-class ConformerChunkEncoder(torch.nn.Module):
+class ConformerChunkEncoder(AbsEncoder):
"""Encoder module definition.
Args:
input_size: Input size.
@@ -1007,7 +1007,7 @@
output_size,
)
- self.output_size = output_size
+ self._output_size = output_size
self.dynamic_chunk_training = dynamic_chunk_training
self.short_chunk_threshold = short_chunk_threshold
@@ -1019,6 +1019,9 @@
self.jitter_range = jitter_range
self.time_reduction_factor = time_reduction_factor
+
+ def output_size(self) -> int:
+ return self._output_size
def get_encoder_input_raw_size(self, size: int, hop_length: int) -> int:
"""Return the corresponding number of sample for a given chunk size, in frames.
@@ -1075,7 +1078,7 @@
limit_size,
)
- mask = make_source_mask(x_len)
+ mask = make_source_mask(x_len).to(x.device)
if self.unified_model_training:
chunk_size = self.default_chunk_size + torch.randint(-self.jitter_range, self.jitter_range+1, (1,)).item()
@@ -1142,7 +1145,7 @@
x = x[:,::self.time_reduction_factor,:]
olens = torch.floor_divide(olens-1, self.time_reduction_factor) + 1
- return x, olens
+ return x, olens, None
def simu_chunk_forward(
self,
--
Gitblit v1.9.1