From 5fec3c9e58fceda85fa2daf7deec2492372dac8a Mon Sep 17 00:00:00 2001
From: Chong Zhang <iriszhangchong@gmail.com>
Date: 星期二, 23 五月 2023 17:01:47 +0800
Subject: [PATCH] Update modelscope_models.md
---
funasr/models/encoder/sanm_encoder.py | 30 +++++++++++++++++++++++-------
1 files changed, 23 insertions(+), 7 deletions(-)
diff --git a/funasr/models/encoder/sanm_encoder.py b/funasr/models/encoder/sanm_encoder.py
index 2a3a353..da67586 100644
--- a/funasr/models/encoder/sanm_encoder.py
+++ b/funasr/models/encoder/sanm_encoder.py
@@ -6,12 +6,14 @@
import logging
import torch
import torch.nn as nn
+import torch.nn.functional as F
from funasr.modules.streaming_utils.chunk_utilis import overlap_chunk
from typeguard import check_argument_types
import numpy as np
+from funasr.torch_utils.device_funcs import to_device
from funasr.modules.nets_utils import make_pad_mask
from funasr.modules.attention import MultiHeadedAttention, MultiHeadedAttentionSANM, MultiHeadedAttentionSANMwithMask
-from funasr.modules.embedding import SinusoidalPositionEncoder
+from funasr.modules.embedding import SinusoidalPositionEncoder, StreamSinusoidalPositionEncoder
from funasr.modules.layer_norm import LayerNorm
from funasr.modules.multi_layer_conv import Conv1dLinear
from funasr.modules.multi_layer_conv import MultiLayeredConv1d
@@ -25,9 +27,10 @@
from funasr.modules.subsampling import Conv2dSubsampling8
from funasr.modules.subsampling import TooShortUttError
from funasr.modules.subsampling import check_short_utt
+from funasr.modules.mask import subsequent_mask, vad_mask
+
from funasr.models.ctc import CTC
from funasr.models.encoder.abs_encoder import AbsEncoder
-from funasr.modules.mask import subsequent_mask, vad_mask
class EncoderLayerSANM(nn.Module):
def __init__(
@@ -117,7 +120,7 @@
class SANMEncoder(AbsEncoder):
"""
- author: Speech Lab, Alibaba Group, China
+ Author: Speech Lab of DAMO Academy, Alibaba Group
San-m: Memory equipped self-attention for end-to-end speech recognition
https://arxiv.org/abs/2006.01713
@@ -180,6 +183,8 @@
self.embed = torch.nn.Linear(input_size, output_size)
elif input_layer == "pe":
self.embed = SinusoidalPositionEncoder()
+ elif input_layer == "pe_online":
+ self.embed = StreamSinusoidalPositionEncoder()
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
@@ -347,6 +352,14 @@
return (xs_pad, intermediate_outs), olens, None
return xs_pad, olens, None
+ def _add_overlap_chunk(self, feats: np.ndarray, cache: dict = {}):
+ if len(cache) == 0:
+ return feats
+ cache["feats"] = to_device(cache["feats"], device=feats.device)
+ overlap_feats = torch.cat((cache["feats"], feats), dim=1)
+ cache["feats"] = overlap_feats[:, -(cache["chunk_size"][0] + cache["chunk_size"][2]):, :]
+ return overlap_feats
+
def forward_chunk(self,
xs_pad: torch.Tensor,
ilens: torch.Tensor,
@@ -357,8 +370,11 @@
if self.embed is None:
xs_pad = xs_pad
else:
- xs_pad = self.embed.forward_chunk(xs_pad, cache)
-
+ xs_pad = self.embed(xs_pad, cache)
+ if cache["tail_chunk"]:
+ xs_pad = to_device(cache["feats"], device=xs_pad.device)
+ else:
+ xs_pad = self._add_overlap_chunk(xs_pad, cache)
encoder_outs = self.encoders0(xs_pad, None, None, None, None)
xs_pad, masks = encoder_outs[0], encoder_outs[1]
intermediate_outs = []
@@ -549,7 +565,7 @@
class SANMEncoderChunkOpt(AbsEncoder):
"""
- 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
@@ -962,7 +978,7 @@
class SANMVadEncoder(AbsEncoder):
"""
- author: Speech Lab, Alibaba Group, China
+ Author: Speech Lab of DAMO Academy, Alibaba Group
"""
--
Gitblit v1.9.1