modelscope/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR.git synced 2026-10-12 02:19:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

shfit to shift (#2266)

Browse Source
This commit is contained in:
Rin Arakaki
2024-12-24 17:51:31 +08:00
committed by GitHub
parent ae7aff2e9c
commit 1367973f98
34 changed files with 216 additions and 216 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/tutorial/Tables.md
+27 -27
View File
@@ -5,13 +5,13 @@
The original intention of the funasr-1.x.x version is to make model integration easier. The core feature is the registry and AutoModel:
* The introduction of the registry enables the development of building blocks to access the model, compatible with a variety of tasks;
* The newly designed AutoModel interface unifies modelscope, huggingface, and funasr inference and training interfaces, and supports free download of repositories;
* Support model export, demo-level service deployment, and industrial-level multi-concurrent service deployment;
* Unify academic and industrial model inference training scripts;
# Quick to get started
@@ -51,19 +51,19 @@ Model = AutoModel(model=[str], device=[str], ncpu=[int], output_dir=[str], batch
```
* `model`(str): [Model Warehouse](https://github.com/alibaba-damo-academy/FunASR/tree/main/model_zoo)The model name in, or the model path in the local disk
* `device`(str): `cuda:0`(Default gpu0), using GPU for inference, specified. If`cpu`Then the CPU is used for inference
* `ncpu`(int): `4`(Default), set the number of threads used for CPU internal operation parallelism
* `output_dir`(str): `None`(Default) If set, the output path of the output result
* `batch_size`(int): `1`(Default), batch processing during decoding, number of samples
* `hub`(str)：`ms`(Default) to download the model from modelscope. If`hf`To download the model from huggingface.
* `**kwargs`(dict): All in`config.yaml`Parameters, which can be specified directly here, for example, the maximum cut length in the vad model.`max_single_segment_time=6000`(Milliseconds).
#### AutoModel reasoning
@@ -72,13 +72,13 @@ Res = model.generate(input=[str], output_dir=[str])
```
* * wav file path, for example: asr\_example.wav
* pcm file path, for example: asr\_example.pcm, you need to specify the audio sampling rate fs (default is 16000)
* Audio byte stream, for example: microphone byte data
* wav.scp,kaldi-style wav list (`wav_id \t wav_path`), for example:
```plaintext
Asr_example1./audios/asr_example1.wav
@@ -89,13 +89,13 @@ Asr_example2./audios/asr_example2.wav
In this input
* Audio sampling points, for example:`audio, rate = soundfile.read("asr_example_zh.wav")`Is numpy.ndarray. batch input is supported. The type is list:`[audio_sample1, audio_sample2, ..., audio_sampleN]`
* fbank input, support group batch. shape is \[batch, frames, dim\], type is torch.Tensor, for example
* `output_dir`: None (default), if set, the output path of the output result
* `**kwargs`(dict): Model-related inference parameters, e.g,`beam_size=10`,`decoding_ctc_weight=0.1`.
Detailed documentation link:[https://github.com/modelscope/FunASR/blob/main/examples/README\_zh.md](https://github.com/modelscope/FunASR/blob/main/examples/README_zh.md)
@@ -128,7 +128,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
@@ -208,10 +208,10 @@ Path resolution: configuration.json (not required)
"model": {"type" : "funasr"},
"pipeline": {"type":"funasr-pipeline"},
"model_name_in_hub": {
"ms":"",
"ms":"",
"hf":""},
"file_path_metas": {
"init_param":"model.pt",
"init_param":"model.pt",
"config":"config.yaml",
"tokenizer_conf": {"bpemodel": "chn_jpn_yue_eng_ko_spectok.bpe.model"},
"frontend_conf":{"cmvn_file": "am.mvn"}}
@@ -274,7 +274,7 @@ class SenseVoiceSmall(nn.Module):
def forward(
self,
**kwargs,
):
):
def inference(
self,
@@ -320,9 +320,9 @@ from funasr.models.sense_voice.model import *
## Principles of Registration
* Model: models are independent of each other. Each Model needs to create a new Model directory under funasr/models/. Do not use class inheritance method!!! Do not import from other model directories, and put everything you need into your own model directory!!! Do not modify the existing model code!!!
* dataset,frontend,tokenizer, if you can reuse the existing one, reuse it directly, if you cannot reuse it, please register a new one, modify it again, and do not modify the original one!!!
# Independent warehouse
@@ -337,7 +337,7 @@ from funasr import AutoModel
model = AutoModel (
model="iic/SenseVoiceSmall ",
trust_remote_code=True
remote_code = "./model.py",
remote_code = "./model.py",
)
```
@@ -360,4 +360,4 @@ res = m.inference(
print(text)
```
Trim reference:[https://github.com/FunAudioLLM/SenseVoice/blob/main/finetune.sh](https://github.com/FunAudioLLM/SenseVoice/blob/main/finetune.sh)
Trim reference:[https://github.com/FunAudioLLM/SenseVoice/blob/main/finetune.sh](https://github.com/FunAudioLLM/SenseVoice/blob/main/finetune.sh)
docs/tutorial/Tables_zh.md
+28 -28
View File
@@ -5,13 +5,13 @@
funasr-1.x.x 版本的设计初衷是【**让模型集成更简单**】，核心feature为注册表与AutoModel：
* 注册表的引入，使得开发中可以用搭积木的方式接入模型，兼容多种task；
* 新设计的AutoModel接口，统一modelscope、huggingface与funasr推理与训练接口，支持自由选择下载仓库；
* 支持模型导出，demo级别服务部署，以及工业级别多并发服务部署；
* 统一学术与工业模型推理训练脚本；
# 快速上手
@@ -51,19 +51,19 @@ model = AutoModel(model=[str], device=[str], ncpu=[int], output_dir=[str], batch
```
* `model`(str): [模型仓库](https://github.com/alibaba-damo-academy/FunASR/tree/main/model_zoo) 中的模型名称，或本地磁盘中的模型路径
* `device`(str): `cuda:0`（默认gpu0），使用 GPU 进行推理，指定。如果为`cpu`，则使用 CPU 进行推理
* `ncpu`(int): `4` （默认），设置用于 CPU 内部操作并行性的线程数
* `output_dir`(str): `None` （默认），如果设置，输出结果的输出路径
* `batch_size`(int): `1` （默认），解码时的批处理，样本个数
* `hub`(str)：`ms`（默认），从modelscope下载模型。如果为`hf`，从huggingface下载模型。
* `**kwargs`(dict): 所有在`config.yaml`中参数，均可以直接在此处指定，例如，vad模型中最大切割长度 `max_single_segment_time=6000` （毫秒）。
#### AutoModel 推理
@@ -72,13 +72,13 @@ res = model.generate(input=[str], output_dir=[str])
```
* * wav文件路径, 例如: asr\_example.wav
* pcm文件路径, 例如: asr\_example.pcm，此时需要指定音频采样率fs（默认为16000）
* 音频字节数流，例如：麦克风的字节数数据
* wav.scp，kaldi 样式的 wav 列表 (`wav_id \t wav_path`), 例如:
```plaintext
asr_example1 ./audios/asr_example1.wav
@@ -89,13 +89,13 @@ asr_example2 ./audios/asr_example2.wav
在这种输入 
* 音频采样点，例如：`audio, rate = soundfile.read("asr_example_zh.wav")`, 数据类型为 numpy.ndarray。支持batch输入，类型为list： `[audio_sample1, audio_sample2, ..., audio_sampleN]`
* fbank输入，支持组batch。shape为\[batch, frames, dim\]，类型为torch.Tensor，例如
* `output_dir`: None （默认），如果设置，输出结果的输出路径
* `**kwargs`(dict): 与模型相关的推理参数，例如，`beam_size=10`，`decoding_ctc_weight=0.1`。
详细文档链接：[https://github.com/modelscope/FunASR/blob/main/examples/README\_zh.md](https://github.com/modelscope/FunASR/blob/main/examples/README_zh.md)
@@ -128,7 +128,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
@@ -208,10 +208,10 @@ scheduler_conf:
"model": {"type" : "funasr"},
"pipeline": {"type":"funasr-pipeline"},
"model_name_in_hub": {
"ms":"",
"ms":"",
"hf":""},
"file_path_metas": {
"init_param":"model.pt",
"init_param":"model.pt",
"config":"config.yaml",
"tokenizer_conf": {"bpemodel": "chn_jpn_yue_eng_ko_spectok.bpe.model"},
"frontend_conf":{"cmvn_file": "am.mvn"}}
@@ -274,7 +274,7 @@ class SenseVoiceSmall(nn.Module):
def forward(
self,
**kwargs,
):
):
def inference(
self,
@@ -320,9 +320,9 @@ from funasr.models.sense_voice.model import *
## 注册原则
* Model：模型之间互相独立，每一个模型，都需要在funasr/models/下面新建一个模型目录，不要采用类的继承方法！！！不要从其他模型目录中import，所有需要用到的都单独放到自己的模型目录中！！！不要修改现有的模型代码！！！
* dataset，frontend，tokenizer，如果能复用现有的，直接复用，如果不能复用，请注册一个新的，再修改，不要修改原来的！！！
# 独立仓库
@@ -336,8 +336,8 @@ from funasr import AutoModel
# trust_remote_code：`True` 表示 model 代码实现从 `remote_code` 处加载，`remote_code` 指定 `model` 具体代码的位置（例如，当前目录下的 `model.py`），支持绝对路径与相对路径，以及网络 url。
model = AutoModel(
model="iic/SenseVoiceSmall",
trust_remote_code=True,
remote_code="./model.py",
trust_remote_code=True,
remote_code="./model.py",
)
```
@@ -360,4 +360,4 @@ res = m.inference(
print(text)
```
微调参考：[https://github.com/FunAudioLLM/SenseVoice/blob/main/finetune.sh](https://github.com/FunAudioLLM/SenseVoice/blob/main/finetune.sh)
微调参考：[https://github.com/FunAudioLLM/SenseVoice/blob/main/finetune.sh](https://github.com/FunAudioLLM/SenseVoice/blob/main/finetune.sh)
examples/industrial_data_pretraining/sanm_kws/conf/sanm_6e_320_256_fdim40_t2602.yaml
+1 -1
View File
@@ -18,7 +18,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
# frontend related
examples/industrial_data_pretraining/sanm_kws_streaming/conf/sanm_6e_320_256_fdim40_t2602.yaml
+1 -1
View File
@@ -18,7 +18,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
chunk_size:
- 16
funasr/models/bicif_paraformer/template.yaml
+2 -2
View File
@@ -30,7 +30,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
# decoder
@@ -45,7 +45,7 @@ decoder_conf:
src_attention_dropout_rate: 0.1
att_layer_num: 16
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
predictor: CifPredictorV3
predictor_conf:
funasr/models/contextual_paraformer/decoder.py
+6 -6
View File
@@ -137,7 +137,7 @@ class ContextualParaformerDecoder(ParaformerSANMDecoder):
concat_after: bool = False,
att_layer_num: int = 6,
kernel_size: int = 21,
sanm_shfit: int = 0,
sanm_shift: int = 0,
):
super().__init__(
vocab_size=vocab_size,
@@ -179,14 +179,14 @@ class ContextualParaformerDecoder(ParaformerSANMDecoder):
self.att_layer_num = att_layer_num
self.num_blocks = num_blocks
if sanm_shfit is None:
sanm_shfit = (kernel_size - 1) // 2
if sanm_shift is None:
sanm_shift = (kernel_size - 1) // 2
self.decoders = repeat(
att_layer_num - 1,
lambda lnum: DecoderLayerSANM(
attention_dim,
MultiHeadedAttentionSANMDecoder(
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=sanm_shift
),
MultiHeadedAttentionCrossAtt(
attention_heads, attention_dim, src_attention_dropout_rate
@@ -210,7 +210,7 @@ class ContextualParaformerDecoder(ParaformerSANMDecoder):
self.last_decoder = ContextualDecoderLayer(
attention_dim,
MultiHeadedAttentionSANMDecoder(
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=sanm_shift
),
MultiHeadedAttentionCrossAtt(
attention_heads, attention_dim, src_attention_dropout_rate
@@ -228,7 +228,7 @@ class ContextualParaformerDecoder(ParaformerSANMDecoder):
lambda lnum: DecoderLayerSANM(
attention_dim,
MultiHeadedAttentionSANMDecoder(
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=0
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=0
),
None,
PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
funasr/models/contextual_paraformer/template.yaml
+3 -3
View File
@@ -30,7 +30,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
@@ -46,7 +46,7 @@ decoder_conf:
src_attention_dropout_rate: 0.1
att_layer_num: 16
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
predictor: CifPredictorV2
predictor_conf:
@@ -126,4 +126,4 @@ ctc_conf:
ctc_type: builtin
reduce: true
ignore_nan_grad: true
normalize: null
normalize: null
funasr/models/ct_transformer/template.yaml
+1 -1
View File
@@ -41,7 +41,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
padding_idx: 0
funasr/models/ct_transformer_streaming/attention.py
+2 -2
View File
@@ -11,9 +11,9 @@ class MultiHeadedAttentionSANMwithMask(MultiHeadedAttentionSANM):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def forward(self, x, mask, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
def forward(self, x, mask, mask_shift_chunk=None, mask_att_chunk_encoder=None):
q_h, k_h, v_h, v = self.forward_qkv(x)
fsmn_memory = self.forward_fsmn(v, mask[0], mask_shfit_chunk)
fsmn_memory = self.forward_fsmn(v, mask[0], mask_shift_chunk)
q_h = q_h * self.d_k ** (-0.5)
scores = torch.matmul(q_h, k_h.transpose(-2, -1))
att_outs = self.forward_attention(v_h, scores, mask[1], mask_att_chunk_encoder)
funasr/models/ct_transformer_streaming/encoder.py
+8 -8
View File
@@ -56,7 +56,7 @@ class EncoderLayerSANM(torch.nn.Module):
self.stochastic_depth_rate = stochastic_depth_rate
self.dropout_rate = dropout_rate
def forward(self, x, mask, cache=None, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
def forward(self, x, mask, cache=None, mask_shift_chunk=None, mask_att_chunk_encoder=None):
"""Compute encoded features.
Args:
@@ -93,7 +93,7 @@ class EncoderLayerSANM(torch.nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
),
),
@@ -109,7 +109,7 @@ class EncoderLayerSANM(torch.nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
)
)
@@ -118,7 +118,7 @@ class EncoderLayerSANM(torch.nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
)
)
@@ -132,7 +132,7 @@ class EncoderLayerSANM(torch.nn.Module):
if not self.normalize_before:
x = self.norm2(x)
return x, mask, cache, mask_shfit_chunk, mask_att_chunk_encoder
return x, mask, cache, mask_shift_chunk, mask_att_chunk_encoder
def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
"""Compute encoded features.
@@ -198,7 +198,7 @@ class SANMVadEncoder(torch.nn.Module):
interctc_layer_idx: List[int] = [],
interctc_use_conditioning: bool = False,
kernel_size: int = 11,
sanm_shfit: int = 0,
sanm_shift: int = 0,
selfattention_layer_type: str = "sanm",
):
super().__init__()
@@ -277,7 +277,7 @@ class SANMVadEncoder(torch.nn.Module):
output_size,
attention_dropout_rate,
kernel_size,
sanm_shfit,
sanm_shift,
)
encoder_selfattn_layer_args = (
@@ -286,7 +286,7 @@ class SANMVadEncoder(torch.nn.Module):
output_size,
attention_dropout_rate,
kernel_size,
sanm_shfit,
sanm_shift,
)
self.encoders0 = repeat(
funasr/models/ct_transformer_streaming/template.yaml
+2 -2
View File
@@ -41,10 +41,10 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 5
sanm_shift: 5
selfattention_layer_type: sanm
padding_idx: 0
tokenizer: CharTokenizer
tokenizer_conf:
unk_symbol: <unk>
unk_symbol: <unk>
funasr/models/monotonic_aligner/template.yaml
+2 -2
View File
@@ -25,7 +25,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
predictor: CifPredictorV3
@@ -111,5 +111,5 @@ ctc_conf:
ctc_type: builtin
reduce: true
ignore_nan_grad: true
normalize: null
funasr/models/paraformer/decoder.py
+9 -9
View File
@@ -248,7 +248,7 @@ class ParaformerSANMDecoder(BaseTransformerDecoder):
concat_after: bool = False,
att_layer_num: int = 6,
kernel_size: int = 21,
sanm_shfit: int = 0,
sanm_shift: int = 0,
lora_list: List[str] = None,
lora_rank: int = 8,
lora_alpha: int = 16,
@@ -298,14 +298,14 @@ class ParaformerSANMDecoder(BaseTransformerDecoder):
self.att_layer_num = att_layer_num
self.num_blocks = num_blocks
if sanm_shfit is None:
sanm_shfit = (kernel_size - 1) // 2
if sanm_shift is None:
sanm_shift = (kernel_size - 1) // 2
self.decoders = repeat(
att_layer_num,
lambda lnum: DecoderLayerSANM(
attention_dim,
MultiHeadedAttentionSANMDecoder(
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=sanm_shift
),
MultiHeadedAttentionCrossAtt(
attention_heads,
@@ -330,7 +330,7 @@ class ParaformerSANMDecoder(BaseTransformerDecoder):
lambda lnum: DecoderLayerSANM(
attention_dim,
MultiHeadedAttentionSANMDecoder(
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=0
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=0
),
None,
PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
@@ -785,20 +785,20 @@ class ParaformerSANMDecoderExport(torch.nn.Module):
for _ in range(cache_num)
]
return (tgt, memory, pre_acoustic_embeds, cache)
def is_optimizable(self):
return True
def get_input_names(self):
cache_num = len(self.model.decoders) + len(self.model.decoders2)
return ['tgt', 'memory', 'pre_acoustic_embeds'] \
+ ['cache_%d' % i for i in range(cache_num)]
def get_output_names(self):
cache_num = len(self.model.decoders) + len(self.model.decoders2)
return ['y'] \
+ ['out_cache_%d' % i for i in range(cache_num)]
def get_dynamic_axes(self):
ret = {
'tgt': {
funasr/models/paraformer/template.yaml
+2 -2
View File
@@ -29,7 +29,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
# decoder
@@ -44,7 +44,7 @@ decoder_conf:
src_attention_dropout_rate: 0.1
att_layer_num: 16
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
predictor: CifPredictorV2
predictor_conf:
funasr/models/paraformer_streaming/model.py
+4 -4
View File
@@ -198,10 +198,10 @@ class ParaformerStreaming(Paraformer):
mask_chunk_predictor = self.encoder.overlap_chunk_cls.get_mask_chunk_predictor(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
mask_shfit_chunk = self.encoder.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.encoder.overlap_chunk_cls.get_mask_shift_chunk(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
encoder_out = encoder_out * mask_shfit_chunk
encoder_out = encoder_out * mask_shift_chunk
pre_acoustic_embeds, pre_token_length, pre_alphas, _ = self.predictor(
encoder_out,
ys_pad,
@@ -357,10 +357,10 @@ class ParaformerStreaming(Paraformer):
mask_chunk_predictor = self.encoder.overlap_chunk_cls.get_mask_chunk_predictor(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
mask_shfit_chunk = self.encoder.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.encoder.overlap_chunk_cls.get_mask_shift_chunk(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
encoder_out = encoder_out * mask_shfit_chunk
encoder_out = encoder_out * mask_shift_chunk
pre_acoustic_embeds, pre_token_length, pre_alphas, pre_peak_index = self.predictor(
encoder_out,
None,
funasr/models/paraformer_streaming/template.yaml
+2 -2
View File
@@ -29,7 +29,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
chunk_size:
- 12
@@ -59,7 +59,7 @@ decoder_conf:
src_attention_dropout_rate: 0.1
att_layer_num: 16
kernel_size: 11
sanm_shfit: 5
sanm_shift: 5
predictor: CifPredictorV2
predictor_conf:
funasr/models/sanm/attention.py
+15 -15
View File
@@ -154,7 +154,7 @@ class MultiHeadedAttentionSANM(nn.Module):
n_feat,
dropout_rate,
kernel_size,
sanm_shfit=0,
sanm_shift=0,
lora_list=None,
lora_rank=8,
lora_alpha=16,
@@ -199,17 +199,17 @@ class MultiHeadedAttentionSANM(nn.Module):
)
# padding
left_padding = (kernel_size - 1) // 2
if sanm_shfit > 0:
left_padding = left_padding + sanm_shfit
if sanm_shift > 0:
left_padding = left_padding + sanm_shift
right_padding = kernel_size - 1 - left_padding
self.pad_fn = nn.ConstantPad1d((left_padding, right_padding), 0.0)
def forward_fsmn(self, inputs, mask, mask_shfit_chunk=None):
def forward_fsmn(self, inputs, mask, mask_shift_chunk=None):
b, t, d = inputs.size()
if mask is not None:
mask = torch.reshape(mask, (b, -1, 1))
if mask_shfit_chunk is not None:
mask = mask * mask_shfit_chunk
if mask_shift_chunk is not None:
mask = mask * mask_shift_chunk
inputs = inputs * mask
x = inputs.transpose(1, 2)
@@ -289,7 +289,7 @@ class MultiHeadedAttentionSANM(nn.Module):
return self.linear_out(x) # (batch, time1, d_model)
def forward(self, x, mask, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
def forward(self, x, mask, mask_shift_chunk=None, mask_att_chunk_encoder=None):
"""Compute scaled dot product attention.
Args:
@@ -304,7 +304,7 @@ class MultiHeadedAttentionSANM(nn.Module):
"""
q_h, k_h, v_h, v = self.forward_qkv(x)
fsmn_memory = self.forward_fsmn(v, mask, mask_shfit_chunk)
fsmn_memory = self.forward_fsmn(v, mask, mask_shift_chunk)
q_h = q_h * self.d_k ** (-0.5)
scores = torch.matmul(q_h, k_h.transpose(-2, -1))
att_outs = self.forward_attention(v_h, scores, mask, mask_att_chunk_encoder)
@@ -478,7 +478,7 @@ class MultiHeadedAttentionSANMDecoder(nn.Module):
"""
def __init__(self, n_feat, dropout_rate, kernel_size, sanm_shfit=0):
def __init__(self, n_feat, dropout_rate, kernel_size, sanm_shift=0):
"""Construct an MultiHeadedAttention object."""
super().__init__()
@@ -490,13 +490,13 @@ class MultiHeadedAttentionSANMDecoder(nn.Module):
# padding
# padding
left_padding = (kernel_size - 1) // 2
if sanm_shfit > 0:
left_padding = left_padding + sanm_shfit
if sanm_shift > 0:
left_padding = left_padding + sanm_shift
right_padding = kernel_size - 1 - left_padding
self.pad_fn = nn.ConstantPad1d((left_padding, right_padding), 0.0)
self.kernel_size = kernel_size
def forward(self, inputs, mask, cache=None, mask_shfit_chunk=None):
def forward(self, inputs, mask, cache=None, mask_shift_chunk=None):
"""
:param x: (#batch, time1, size).
:param mask: Mask tensor (#batch, 1, time)
@@ -509,9 +509,9 @@ class MultiHeadedAttentionSANMDecoder(nn.Module):
if mask is not None:
mask = torch.reshape(mask, (b, -1, 1))
# logging.info("in fsmn, mask: {}, {}".format(mask.size(), mask[0:100:50, :, :]))
if mask_shfit_chunk is not None:
# logging.info("in fsmn, mask_fsmn: {}, {}".format(mask_shfit_chunk.size(), mask_shfit_chunk[0:100:50, :, :]))
mask = mask * mask_shfit_chunk
if mask_shift_chunk is not None:
# logging.info("in fsmn, mask_fsmn: {}, {}".format(mask_shift_chunk.size(), mask_shift_chunk[0:100:50, :, :]))
mask = mask * mask_shift_chunk
# logging.info("in fsmn, mask_after_fsmn: {}, {}".format(mask.size(), mask[0:100:50, :, :]))
# print("in fsmn, mask", mask.size())
# print("in fsmn, inputs", inputs.size())
funasr/models/sanm/decoder.py
+5 -5
View File
@@ -226,7 +226,7 @@ class FsmnDecoder(BaseTransformerDecoder):
concat_after: bool = False,
att_layer_num: int = 6,
kernel_size: int = 21,
sanm_shfit: int = None,
sanm_shift: int = None,
concat_embeds: bool = False,
attention_dim: int = None,
tf2torch_tensor_name_prefix_torch: str = "decoder",
@@ -271,14 +271,14 @@ class FsmnDecoder(BaseTransformerDecoder):
self.att_layer_num = att_layer_num
self.num_blocks = num_blocks
if sanm_shfit is None:
sanm_shfit = (kernel_size - 1) // 2
if sanm_shift is None:
sanm_shift = (kernel_size - 1) // 2
self.decoders = repeat(
att_layer_num,
lambda lnum: DecoderLayerSANM(
attention_dim,
MultiHeadedAttentionSANMDecoder(
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=sanm_shift
),
MultiHeadedAttentionCrossAtt(
attention_heads,
@@ -303,7 +303,7 @@ class FsmnDecoder(BaseTransformerDecoder):
attention_dim,
self_attention_dropout_rate,
kernel_size,
sanm_shfit=sanm_shfit,
sanm_shift=sanm_shift,
),
None,
PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
funasr/models/sanm/encoder.py
+8 -8
View File
@@ -69,7 +69,7 @@ class EncoderLayerSANM(nn.Module):
self.stochastic_depth_rate = stochastic_depth_rate
self.dropout_rate = dropout_rate
def forward(self, x, mask, cache=None, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
def forward(self, x, mask, cache=None, mask_shift_chunk=None, mask_att_chunk_encoder=None):
"""Compute encoded features.
Args:
@@ -106,7 +106,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
),
),
@@ -122,7 +122,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
)
)
@@ -131,7 +131,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
)
)
@@ -145,7 +145,7 @@ class EncoderLayerSANM(nn.Module):
if not self.normalize_before:
x = self.norm2(x)
return x, mask, cache, mask_shfit_chunk, mask_att_chunk_encoder
return x, mask, cache, mask_shift_chunk, mask_att_chunk_encoder
def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
"""Compute encoded features.
@@ -212,7 +212,7 @@ class SANMEncoder(nn.Module):
interctc_layer_idx: List[int] = [],
interctc_use_conditioning: bool = False,
kernel_size: int = 11,
sanm_shfit: int = 0,
sanm_shift: int = 0,
lora_list: List[str] = None,
lora_rank: int = 8,
lora_alpha: int = 16,
@@ -299,7 +299,7 @@ class SANMEncoder(nn.Module):
output_size,
attention_dropout_rate,
kernel_size,
sanm_shfit,
sanm_shift,
lora_list,
lora_rank,
lora_alpha,
@@ -312,7 +312,7 @@ class SANMEncoder(nn.Module):
output_size,
attention_dropout_rate,
kernel_size,
sanm_shfit,
sanm_shift,
lora_list,
lora_rank,
lora_alpha,
funasr/models/sanm/template.yaml
+2 -2
View File
@@ -26,7 +26,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
# decoder
@@ -41,7 +41,7 @@ decoder_conf:
src_attention_dropout_rate: 0.1
att_layer_num: 16
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
funasr/models/scama/chunk_utilis.py
+20 -20
View File
@@ -21,7 +21,7 @@ class overlap_chunk:
stride: tuple = (10,),
pad_left: tuple = (0,),
encoder_att_look_back_factor: tuple = (1,),
shfit_fsmn: int = 0,
shift_fsmn: int = 0,
decoder_att_look_back_factor: tuple = (1,),
):
@@ -45,11 +45,11 @@ class overlap_chunk:
encoder_att_look_back_factor,
decoder_att_look_back_factor,
)
self.shfit_fsmn = shfit_fsmn
self.shift_fsmn = shift_fsmn
self.x_add_mask = None
self.x_rm_mask = None
self.x_len = None
self.mask_shfit_chunk = None
self.mask_shift_chunk = None
self.mask_chunk_predictor = None
self.mask_att_chunk_encoder = None
self.mask_shift_att_chunk_decoder = None
@@ -88,7 +88,7 @@ class overlap_chunk:
stride,
pad_left,
encoder_att_look_back_factor,
chunk_size + self.shfit_fsmn,
chunk_size + self.shift_fsmn,
decoder_att_look_back_factor,
)
return (
@@ -118,13 +118,13 @@ class overlap_chunk:
chunk_size, stride, pad_left, encoder_att_look_back_factor, chunk_size_pad_shift = (
self.get_chunk_size(ind)
)
shfit_fsmn = self.shfit_fsmn
shift_fsmn = self.shift_fsmn
pad_right = chunk_size - stride - pad_left
chunk_num_batch = np.ceil(x_len / stride).astype(np.int32)
x_len_chunk = (
(chunk_num_batch - 1) * chunk_size_pad_shift
+ shfit_fsmn
+ shift_fsmn
+ pad_left
+ 0
+ x_len
@@ -138,13 +138,13 @@ class overlap_chunk:
max_len_for_x_mask_tmp = max(chunk_size, x_len_max + pad_left)
x_add_mask = np.zeros([0, max_len_for_x_mask_tmp], dtype=dtype)
x_rm_mask = np.zeros([max_len_for_x_mask_tmp, 0], dtype=dtype)
mask_shfit_chunk = np.zeros([0, num_units], dtype=dtype)
mask_shift_chunk = np.zeros([0, num_units], dtype=dtype)
mask_chunk_predictor = np.zeros([0, num_units_predictor], dtype=dtype)
mask_shift_att_chunk_decoder = np.zeros([0, 1], dtype=dtype)
mask_att_chunk_encoder = np.zeros([0, chunk_num * chunk_size_pad_shift], dtype=dtype)
for chunk_ids in range(chunk_num):
# x_mask add
fsmn_padding = np.zeros((shfit_fsmn, max_len_for_x_mask_tmp), dtype=dtype)
fsmn_padding = np.zeros((shift_fsmn, max_len_for_x_mask_tmp), dtype=dtype)
x_mask_cur = np.diag(np.ones(chunk_size, dtype=np.float32))
x_mask_pad_left = np.zeros((chunk_size, chunk_ids * stride), dtype=dtype)
x_mask_pad_right = np.zeros((chunk_size, max_len_for_x_mask_tmp), dtype=dtype)
@@ -154,7 +154,7 @@ class overlap_chunk:
x_add_mask = np.concatenate([x_add_mask, x_add_mask_fsmn], axis=0)
# x_mask rm
fsmn_padding = np.zeros((max_len_for_x_mask_tmp, shfit_fsmn), dtype=dtype)
fsmn_padding = np.zeros((max_len_for_x_mask_tmp, shift_fsmn), dtype=dtype)
padding_mask_left = np.zeros((max_len_for_x_mask_tmp, pad_left), dtype=dtype)
padding_mask_right = np.zeros((max_len_for_x_mask_tmp, pad_right), dtype=dtype)
x_mask_cur = np.diag(np.ones(stride, dtype=dtype))
@@ -170,13 +170,13 @@ class overlap_chunk:
x_rm_mask = np.concatenate([x_rm_mask, x_rm_mask_cur_fsmn], axis=1)
# fsmn_padding_mask
pad_shfit_mask = np.zeros([shfit_fsmn, num_units], dtype=dtype)
pad_shift_mask = np.zeros([shift_fsmn, num_units], dtype=dtype)
ones_1 = np.ones([chunk_size, num_units], dtype=dtype)
mask_shfit_chunk_cur = np.concatenate([pad_shfit_mask, ones_1], axis=0)
mask_shfit_chunk = np.concatenate([mask_shfit_chunk, mask_shfit_chunk_cur], axis=0)
mask_shift_chunk_cur = np.concatenate([pad_shift_mask, ones_1], axis=0)
mask_shift_chunk = np.concatenate([mask_shift_chunk, mask_shift_chunk_cur], axis=0)
# predictor mask
zeros_1 = np.zeros([shfit_fsmn + pad_left, num_units_predictor], dtype=dtype)
zeros_1 = np.zeros([shift_fsmn + pad_left, num_units_predictor], dtype=dtype)
ones_2 = np.ones([stride, num_units_predictor], dtype=dtype)
zeros_3 = np.zeros(
[chunk_size - stride - pad_left, num_units_predictor], dtype=dtype
@@ -188,13 +188,13 @@ class overlap_chunk:
)
# encoder att mask
zeros_1_top = np.zeros([shfit_fsmn, chunk_num * chunk_size_pad_shift], dtype=dtype)
zeros_1_top = np.zeros([shift_fsmn, chunk_num * chunk_size_pad_shift], dtype=dtype)
zeros_2_num = max(chunk_ids - encoder_att_look_back_factor, 0)
zeros_2 = np.zeros([chunk_size, zeros_2_num * chunk_size_pad_shift], dtype=dtype)
encoder_att_look_back_num = max(chunk_ids - zeros_2_num, 0)
zeros_2_left = np.zeros([chunk_size, shfit_fsmn], dtype=dtype)
zeros_2_left = np.zeros([chunk_size, shift_fsmn], dtype=dtype)
ones_2_mid = np.ones([stride, stride], dtype=dtype)
zeros_2_bottom = np.zeros([chunk_size - stride, stride], dtype=dtype)
zeros_2_right = np.zeros([chunk_size, chunk_size - stride], dtype=dtype)
@@ -202,7 +202,7 @@ class overlap_chunk:
ones_2 = np.concatenate([zeros_2_left, ones_2, zeros_2_right], axis=1)
ones_2 = np.tile(ones_2, [1, encoder_att_look_back_num])
zeros_3_left = np.zeros([chunk_size, shfit_fsmn], dtype=dtype)
zeros_3_left = np.zeros([chunk_size, shift_fsmn], dtype=dtype)
ones_3_right = np.ones([chunk_size, chunk_size], dtype=dtype)
ones_3 = np.concatenate([zeros_3_left, ones_3_right], axis=1)
@@ -218,7 +218,7 @@ class overlap_chunk:
)
# decoder fsmn_shift_att_mask
zeros_1 = np.zeros([shfit_fsmn, 1])
zeros_1 = np.zeros([shift_fsmn, 1])
ones_1 = np.ones([chunk_size, 1])
mask_shift_att_chunk_decoder_cur = np.concatenate([zeros_1, ones_1], axis=0)
mask_shift_att_chunk_decoder = np.concatenate(
@@ -229,7 +229,7 @@ class overlap_chunk:
self.x_len_chunk = x_len_chunk
self.x_rm_mask = x_rm_mask[:x_len_max, :x_len_chunk_max]
self.x_len = x_len
self.mask_shfit_chunk = mask_shfit_chunk[:x_len_chunk_max, :]
self.mask_shift_chunk = mask_shift_chunk[:x_len_chunk_max, :]
self.mask_chunk_predictor = mask_chunk_predictor[:x_len_chunk_max, :]
self.mask_att_chunk_encoder = mask_att_chunk_encoder[:x_len_chunk_max, :x_len_chunk_max]
self.mask_shift_att_chunk_decoder = mask_shift_att_chunk_decoder[:x_len_chunk_max, :]
@@ -238,7 +238,7 @@ class overlap_chunk:
self.x_len_chunk,
self.x_rm_mask,
self.x_len,
self.mask_shfit_chunk,
self.mask_shift_chunk,
self.mask_chunk_predictor,
self.mask_att_chunk_encoder,
self.mask_shift_att_chunk_decoder,
@@ -309,7 +309,7 @@ class overlap_chunk:
x = torch.from_numpy(x).type(dtype).to(device)
return x
def get_mask_shfit_chunk(
def get_mask_shift_chunk(
self, chunk_outs=None, device="cpu", batch_size=1, num_units=1, idx=4, dtype=torch.float32
):
with torch.no_grad():
funasr/models/scama/decoder.py
+5 -5
View File
@@ -226,7 +226,7 @@ class FsmnDecoderSCAMAOpt(BaseTransformerDecoder):
concat_after: bool = False,
att_layer_num: int = 6,
kernel_size: int = 21,
sanm_shfit: int = None,
sanm_shift: int = None,
concat_embeds: bool = False,
attention_dim: int = None,
tf2torch_tensor_name_prefix_torch: str = "decoder",
@@ -271,14 +271,14 @@ class FsmnDecoderSCAMAOpt(BaseTransformerDecoder):
self.att_layer_num = att_layer_num
self.num_blocks = num_blocks
if sanm_shfit is None:
sanm_shfit = (kernel_size - 1) // 2
if sanm_shift is None:
sanm_shift = (kernel_size - 1) // 2
self.decoders = repeat(
att_layer_num,
lambda lnum: DecoderLayerSANM(
attention_dim,
MultiHeadedAttentionSANMDecoder(
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shift=sanm_shift
),
MultiHeadedAttentionCrossAtt(
attention_heads,
@@ -303,7 +303,7 @@ class FsmnDecoderSCAMAOpt(BaseTransformerDecoder):
attention_dim,
self_attention_dropout_rate,
kernel_size,
sanm_shfit=sanm_shfit,
sanm_shift=sanm_shift,
),
None,
PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
funasr/models/scama/encoder.py
+15 -15
View File
@@ -69,7 +69,7 @@ class EncoderLayerSANM(nn.Module):
self.stochastic_depth_rate = stochastic_depth_rate
self.dropout_rate = dropout_rate
def forward(self, x, mask, cache=None, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
def forward(self, x, mask, cache=None, mask_shift_chunk=None, mask_att_chunk_encoder=None):
"""Compute encoded features.
Args:
@@ -106,7 +106,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
),
),
@@ -122,7 +122,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
)
)
@@ -131,7 +131,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
)
)
@@ -145,7 +145,7 @@ class EncoderLayerSANM(nn.Module):
if not self.normalize_before:
x = self.norm2(x)
return x, mask, cache, mask_shfit_chunk, mask_att_chunk_encoder
return x, mask, cache, mask_shift_chunk, mask_att_chunk_encoder
def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
"""Compute encoded features.
@@ -212,7 +212,7 @@ class SANMEncoderChunkOpt(nn.Module):
interctc_layer_idx: List[int] = [],
interctc_use_conditioning: bool = False,
kernel_size: int = 11,
sanm_shfit: int = 0,
sanm_shift: int = 0,
selfattention_layer_type: str = "sanm",
chunk_size: Union[int, Sequence[int]] = (16,),
stride: Union[int, Sequence[int]] = (10,),
@@ -299,7 +299,7 @@ class SANMEncoderChunkOpt(nn.Module):
output_size,
attention_dropout_rate,
kernel_size,
sanm_shfit,
sanm_shift,
)
encoder_selfattn_layer_args = (
@@ -308,7 +308,7 @@ class SANMEncoderChunkOpt(nn.Module):
output_size,
attention_dropout_rate,
kernel_size,
sanm_shfit,
sanm_shift,
)
self.encoders0 = repeat(
1,
@@ -343,12 +343,12 @@ class SANMEncoderChunkOpt(nn.Module):
assert 0 < min(interctc_layer_idx) and max(interctc_layer_idx) < num_blocks
self.interctc_use_conditioning = interctc_use_conditioning
self.conditioning_layer = None
shfit_fsmn = (kernel_size - 1) // 2
shift_fsmn = (kernel_size - 1) // 2
self.overlap_chunk_cls = overlap_chunk(
chunk_size=chunk_size,
stride=stride,
pad_left=pad_left,
shfit_fsmn=shfit_fsmn,
shift_fsmn=shift_fsmn,
encoder_att_look_back_factor=encoder_att_look_back_factor,
decoder_att_look_back_factor=decoder_att_look_back_factor,
)
@@ -397,31 +397,31 @@ class SANMEncoderChunkOpt(nn.Module):
else:
xs_pad = self.embed(xs_pad)
mask_shfit_chunk, mask_att_chunk_encoder = None, None
mask_shift_chunk, mask_att_chunk_encoder = None, None
if self.overlap_chunk_cls is not None:
ilens = masks.squeeze(1).sum(1)
chunk_outs = self.overlap_chunk_cls.gen_chunk_mask(ilens, ind)
xs_pad, ilens = self.overlap_chunk_cls.split_chunk(xs_pad, ilens, chunk_outs=chunk_outs)
masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)
mask_shfit_chunk = self.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.overlap_chunk_cls.get_mask_shift_chunk(
chunk_outs, xs_pad.device, xs_pad.size(0), dtype=xs_pad.dtype
)
mask_att_chunk_encoder = self.overlap_chunk_cls.get_mask_att_chunk_encoder(
chunk_outs, xs_pad.device, xs_pad.size(0), dtype=xs_pad.dtype
)
encoder_outs = self.encoders0(xs_pad, masks, None, mask_shfit_chunk, mask_att_chunk_encoder)
encoder_outs = self.encoders0(xs_pad, masks, None, mask_shift_chunk, mask_att_chunk_encoder)
xs_pad, masks = encoder_outs[0], encoder_outs[1]
intermediate_outs = []
if len(self.interctc_layer_idx) == 0:
encoder_outs = self.encoders(
xs_pad, masks, None, mask_shfit_chunk, mask_att_chunk_encoder
xs_pad, masks, None, mask_shift_chunk, mask_att_chunk_encoder
)
xs_pad, masks = encoder_outs[0], encoder_outs[1]
else:
for layer_idx, encoder_layer in enumerate(self.encoders):
encoder_outs = encoder_layer(
xs_pad, masks, None, mask_shfit_chunk, mask_att_chunk_encoder
xs_pad, masks, None, mask_shift_chunk, mask_att_chunk_encoder
)
xs_pad, masks = encoder_outs[0], encoder_outs[1]
if layer_idx + 1 in self.interctc_layer_idx:
funasr/models/scama/model.py
+4 -4
View File
@@ -321,10 +321,10 @@ class SCAMA(nn.Module):
mask_chunk_predictor = self.encoder.overlap_chunk_cls.get_mask_chunk_predictor(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
mask_shfit_chunk = self.encoder.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.encoder.overlap_chunk_cls.get_mask_shift_chunk(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
encoder_out = encoder_out * mask_shfit_chunk
encoder_out = encoder_out * mask_shift_chunk
pre_acoustic_embeds, pre_token_length, pre_alphas, _ = self.predictor(
encoder_out,
ys_out_pad,
@@ -415,10 +415,10 @@ class SCAMA(nn.Module):
mask_chunk_predictor = self.encoder.overlap_chunk_cls.get_mask_chunk_predictor(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
mask_shfit_chunk = self.encoder.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.encoder.overlap_chunk_cls.get_mask_shift_chunk(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
encoder_out = encoder_out * mask_shfit_chunk
encoder_out = encoder_out * mask_shift_chunk
pre_acoustic_embeds, pre_token_length, pre_alphas, _ = self.predictor(
encoder_out,
ys_out_pad,
funasr/models/scama/template.yaml
+2 -2
View File
@@ -26,7 +26,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
# decoder
@@ -41,7 +41,7 @@ decoder_conf:
src_attention_dropout_rate: 0.1
att_layer_num: 16
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
predictor: CifPredictorV2
predictor_conf:
funasr/models/seaco_paraformer/template.yaml
+3 -3
View File
@@ -36,7 +36,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
# decoder
@@ -51,7 +51,7 @@ decoder_conf:
src_attention_dropout_rate: 0.1
att_layer_num: 16
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
# seaco decoder
seaco_decoder: ParaformerSANMDecoder
@@ -64,7 +64,7 @@ seaco_decoder_conf:
self_attention_dropout_rate: 0.1
src_attention_dropout_rate: 0.1
kernel_size: 21
sanm_shfit: 0
sanm_shift: 0
use_output_layer: false
wo_input_layer: true
funasr/models/sense_voice/model.py
+16 -16
View File
@@ -95,7 +95,7 @@ class MultiHeadedAttentionSANM(nn.Module):
n_feat,
dropout_rate,
kernel_size,
sanm_shfit=0,
sanm_shift=0,
lora_list=None,
lora_rank=8,
lora_alpha=16,
@@ -121,17 +121,17 @@ class MultiHeadedAttentionSANM(nn.Module):
)
# padding
left_padding = (kernel_size - 1) // 2
if sanm_shfit > 0:
left_padding = left_padding + sanm_shfit
if sanm_shift > 0:
left_padding = left_padding + sanm_shift
right_padding = kernel_size - 1 - left_padding
self.pad_fn = nn.ConstantPad1d((left_padding, right_padding), 0.0)
def forward_fsmn(self, inputs, mask, mask_shfit_chunk=None):
def forward_fsmn(self, inputs, mask, mask_shift_chunk=None):
b, t, d = inputs.size()
if mask is not None:
mask = torch.reshape(mask, (b, -1, 1))
if mask_shfit_chunk is not None:
mask = mask * mask_shfit_chunk
if mask_shift_chunk is not None:
mask = mask * mask_shift_chunk
inputs = inputs * mask
x = inputs.transpose(1, 2)
@@ -211,7 +211,7 @@ class MultiHeadedAttentionSANM(nn.Module):
return self.linear_out(x) # (batch, time1, d_model)
def forward(self, x, mask, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
def forward(self, x, mask, mask_shift_chunk=None, mask_att_chunk_encoder=None):
"""Compute scaled dot product attention.
Args:
@@ -226,7 +226,7 @@ class MultiHeadedAttentionSANM(nn.Module):
"""
q_h, k_h, v_h, v = self.forward_qkv(x)
fsmn_memory = self.forward_fsmn(v, mask, mask_shfit_chunk)
fsmn_memory = self.forward_fsmn(v, mask, mask_shift_chunk)
q_h = q_h * self.d_k ** (-0.5)
scores = torch.matmul(q_h, k_h.transpose(-2, -1))
att_outs = self.forward_attention(v_h, scores, mask, mask_att_chunk_encoder)
@@ -326,7 +326,7 @@ class EncoderLayerSANM(nn.Module):
self.stochastic_depth_rate = stochastic_depth_rate
self.dropout_rate = dropout_rate
def forward(self, x, mask, cache=None, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
def forward(self, x, mask, cache=None, mask_shift_chunk=None, mask_att_chunk_encoder=None):
"""Compute encoded features.
Args:
@@ -363,7 +363,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
),
),
@@ -379,7 +379,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
)
)
@@ -388,7 +388,7 @@ class EncoderLayerSANM(nn.Module):
self.self_attn(
x,
mask,
mask_shfit_chunk=mask_shfit_chunk,
mask_shift_chunk=mask_shift_chunk,
mask_att_chunk_encoder=mask_att_chunk_encoder,
)
)
@@ -402,7 +402,7 @@ class EncoderLayerSANM(nn.Module):
if not self.normalize_before:
x = self.norm2(x)
return x, mask, cache, mask_shfit_chunk, mask_att_chunk_encoder
return x, mask, cache, mask_shift_chunk, mask_att_chunk_encoder
def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
"""Compute encoded features.
@@ -469,7 +469,7 @@ class SenseVoiceEncoderSmall(nn.Module):
positionwise_conv_kernel_size: int = 1,
padding_idx: int = -1,
kernel_size: int = 11,
sanm_shfit: int = 0,
sanm_shift: int = 0,
selfattention_layer_type: str = "sanm",
**kwargs,
):
@@ -494,7 +494,7 @@ class SenseVoiceEncoderSmall(nn.Module):
output_size,
attention_dropout_rate,
kernel_size,
sanm_shfit,
sanm_shift,
)
encoder_selfattn_layer_args = (
attention_heads,
@@ -502,7 +502,7 @@ class SenseVoiceEncoderSmall(nn.Module):
output_size,
attention_dropout_rate,
kernel_size,
sanm_shfit,
sanm_shift,
)
self.encoders0 = nn.ModuleList(
funasr/models/sond/encoder/fsmn_encoder.py
+3 -3
View File
@@ -36,12 +36,12 @@ class FsmnBlock(torch.nn.Module):
right_padding = kernel_size - 1 - left_padding
self.pad_fn = nn.ConstantPad1d((left_padding, right_padding), 0.0)
def forward(self, inputs, mask, mask_shfit_chunk=None):
def forward(self, inputs, mask, mask_shift_chunk=None):
b, t, d = inputs.size()
if mask is not None:
mask = torch.reshape(mask, (b, -1, 1))
if mask_shfit_chunk is not None:
mask = mask * mask_shfit_chunk
if mask_shift_chunk is not None:
mask = mask * mask_shift_chunk
inputs = inputs * mask
x = inputs.transpose(1, 2)
funasr/models/uniasr/model.py
+8 -8
View File
@@ -521,10 +521,10 @@ class UniASR(torch.nn.Module):
mask_chunk_predictor = self.encoder.overlap_chunk_cls.get_mask_chunk_predictor(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
mask_shfit_chunk = self.encoder.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.encoder.overlap_chunk_cls.get_mask_shift_chunk(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
encoder_out = encoder_out * mask_shfit_chunk
encoder_out = encoder_out * mask_shift_chunk
pre_acoustic_embeds, pre_token_length, pre_alphas, _ = self.predictor(
encoder_out,
ys_out_pad,
@@ -622,10 +622,10 @@ class UniASR(torch.nn.Module):
mask_chunk_predictor = self.encoder2.overlap_chunk_cls.get_mask_chunk_predictor(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
mask_shfit_chunk = self.encoder2.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.encoder2.overlap_chunk_cls.get_mask_shift_chunk(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
encoder_out = encoder_out * mask_shfit_chunk
encoder_out = encoder_out * mask_shift_chunk
pre_acoustic_embeds, pre_token_length, pre_alphas, _ = self.predictor2(
encoder_out,
ys_out_pad,
@@ -724,10 +724,10 @@ class UniASR(torch.nn.Module):
mask_chunk_predictor = self.encoder.overlap_chunk_cls.get_mask_chunk_predictor(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
mask_shfit_chunk = self.encoder.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.encoder.overlap_chunk_cls.get_mask_shift_chunk(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
encoder_out = encoder_out * mask_shfit_chunk
encoder_out = encoder_out * mask_shift_chunk
pre_acoustic_embeds, pre_token_length, pre_alphas, _ = self.predictor(
encoder_out,
ys_out_pad,
@@ -806,10 +806,10 @@ class UniASR(torch.nn.Module):
mask_chunk_predictor = self.encoder2.overlap_chunk_cls.get_mask_chunk_predictor(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
mask_shfit_chunk = self.encoder2.overlap_chunk_cls.get_mask_shfit_chunk(
mask_shift_chunk = self.encoder2.overlap_chunk_cls.get_mask_shift_chunk(
None, device=encoder_out.device, batch_size=encoder_out.size(0)
)
encoder_out = encoder_out * mask_shfit_chunk
encoder_out = encoder_out * mask_shift_chunk
pre_acoustic_embeds, pre_token_length, pre_alphas, _ = self.predictor2(
encoder_out,
ys_out_pad,
funasr/models/uniasr/template.yaml
+2 -2
View File
@@ -33,7 +33,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
chunk_size:
- 20
@@ -89,7 +89,7 @@ encoder2_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 21
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
chunk_size:
- 45
runtime/csharp/AliParaformerAsr/AliParaformerAsr/Model/DecoderConfEntity.cs
+3 -3
View File
@@ -19,7 +19,7 @@ namespace AliParaformerAsr.Model
private float _src_attention_dropout_rate = 0.1F;
private int _att_layer_num = 16;
private int _kernel_size = 11;
private int _sanm_shfit = 0;
private int _sanm_shift = 0;
public int attention_heads { get => _attention_heads; set => _attention_heads = value; }
public int linear_units { get => _linear_units; set => _linear_units = value; }
@@ -30,7 +30,7 @@ namespace AliParaformerAsr.Model
public float src_attention_dropout_rate { get => _src_attention_dropout_rate; set => _src_attention_dropout_rate = value; }
public int att_layer_num { get => _att_layer_num; set => _att_layer_num = value; }
public int kernel_size { get => _kernel_size; set => _kernel_size = value; }
public int sanm_shfit { get => _sanm_shfit; set => _sanm_shfit = value; }
public int sanm_shift { get => _sanm_shift; set => _sanm_shift = value; }
}
}
runtime/csharp/AliParaformerAsr/AliParaformerAsr/Model/EncoderConfEntity.cs
+2 -2
View File
@@ -21,7 +21,7 @@ namespace AliParaformerAsr.Model
private string _pos_enc_class = "SinusoidalPositionEncoder";
private bool _normalize_before = true;
private int _kernel_size = 11;
private int _sanm_shfit = 0;
private int _sanm_shift = 0;
private string _selfattention_layer_type = "sanm";
public int output_size { get => _output_size; set => _output_size = value; }
@@ -35,7 +35,7 @@ namespace AliParaformerAsr.Model
public string pos_enc_class { get => _pos_enc_class; set => _pos_enc_class = value; }
public bool normalize_before { get => _normalize_before; set => _normalize_before = value; }
public int kernel_size { get => _kernel_size; set => _kernel_size = value; }
public int sanm_shfit { get => _sanm_shfit; set => _sanm_shfit = value; }
public int sanm_shift { get => _sanm_shift; set => _sanm_shift = value; }
public string selfattention_layer_type { get => _selfattention_layer_type; set => _selfattention_layer_type = value; }
}
}
runtime/triton_gpu/model_repo_paraformer_large_online/feature_extractor/config.yaml
+2 -2
View File
@@ -8593,7 +8593,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm
chunk_size:
- 12
@@ -8623,7 +8623,7 @@ decoder_conf:
src_attention_dropout_rate: 0.1
att_layer_num: 16
kernel_size: 11
sanm_shfit: 5
sanm_shift: 5
predictor: cif_predictor_v2
predictor_conf:
idim: 512
runtime/triton_gpu/model_repo_sense_voice_small/feature_extractor/config.yaml
+1 -1
View File
@@ -12,7 +12,7 @@ encoder_conf:
pos_enc_class: SinusoidalPositionEncoder
normalize_before: true
kernel_size: 11
sanm_shfit: 0
sanm_shift: 0
selfattention_layer_type: sanm