EngineX-MetaX/enginex-c_series-asr
9
1
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

first revise

Browse Source
This commit is contained in:
aceforeverd
2025-08-28 19:00:03 +08:00
parent a4ec58a45e
commit 21dbd460b2
5 changed files with 30 additions and 809 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
Dockerfile Normal file
View File

@@ -0,0 +1,12 @@
FROM git.modelhub.org.cn:9443/enginex-metax/maca-c500-pytorch:2.33.0.6-torch2.6-py310-ubuntu24.04-amd64
RUN /opt/conda/bin/pip install funasr modelscope huggingface_hub
RUN chmod 1777 -R /tmp && apt update && apt install -y ffmpeg
WORKDIR /opt/app
COPY ./ ./
RUN /opt/conda/bin/pip install -r requirements.txt
EXPOSE 80
ENTRYPOINT ["python3", "./test_funasr.py"]

20
Dockerfile.funasr-mr100
View File

@@ -1,20 +0,0 @@
FROM git.modelhub.org.cn:9443/enginex-iluvatar/mr100_corex:4.3.0
WORKDIR /root
COPY requirements.txt /root
RUN pip install -r requirements.txt
RUN apt update && apt install -y vim net-tools
RUN pip install funasr==1.2.6 openai-whisper
ADD . /root/
ADD nltk_data.tar.gz /root/
RUN tar -xvzf nltk_data.tar.gz
RUN cp ./replaced_files/mr_v100/cif_predictor.py /usr/local/lib/python3.10/site-packages/funasr/models/paraformer/
EXPOSE 80
ENTRYPOINT ["bash"]
CMD ["./start_funasr.sh"]

17
README.md
View File

@@ -1,8 +1,8 @@
# 天数智芯 智铠100 FunASR
# 沐曦 MetaX C500 FunASR
## 镜像构造
```shell
docker build -f ./Dockerfile.funasr-mr100 -t <your_image> .
docker build -t <built_img> .
```
## 使用说明
@@ -11,9 +11,8 @@ docker build -f ./Dockerfile.funasr-mr100 -t <your_image> .
1. 本项目中附带上了示例测试数据，音频文件为`lei-jun-test.wav`，音频的识别准确内容文件为`lei-jun.txt`，用户需要准备好相应的ASR模型路径，本例中假设我们已经下载好了SenseVoiceSmall模型存放于/model/SenseVoiceSmall
2. 在本项目路径下执行以下快速测试命令
```shell
docker run -it \
-v /usr/src:/usr/src \
-v /lib/modules:/lib/modules --device=/dev/iluvatar0:/dev/iluvatar0 \
metax-docker run -it \
--gpus=[0] \
-v $PWD:/tmp/workspace \
-v /model:/model \
-e MODEL_DIR=/model/SenseVoiceSmall \
@@ -21,7 +20,7 @@ docker run -it \
-e ANSWER_FILE=lei-jun.txt \
-e RESULT_FILE=result.json \
--cpus=4 --memory=16g \
<your_image>
<built_img>
```
上述测试指令成功运行将会在terminal中看到对测试音频的识别结果，运行时间以及1-cer效果指标，并且当前文件下会生成一个`result.json`文件记录刚才的测试结果
@@ -29,8 +28,8 @@ docker run -it \
用户可使用类似上述的docker run指令以交互形式进入镜像中，主要的测试代码为`test_funasr.py`,用户可自行修改代码中需要测试的模型路径、测试文件路径以及调用funASR逻辑
## 智铠100模型适配情况
我们在智铠100上针对funASR部分进行了所有大类的适配，测试方式为在Nvidia A100环境下和智铠100加速卡上对同一段长音频进行语音识别任务，获取运行时间，1-cer指标。运行时都只使用一张显卡
## MetaX C500 模型适配情况
我们在 MetaX C500 上针对funASR部分进行了所有大类的适配，测试方式为在Nvidia A100环境下和智铠100加速卡上对同一段长音频进行语音识别任务，获取运行时间，1-cer指标。运行时都只使用一张显卡
| 模型大类 | 模型地址 |A100运行时间(秒)|智铠100运行时间(秒)|A100 1-cer|智铠100 1-cer| 备注 |
|------|---------------|-----|----|-------|-------|---------------------|
@@ -38,4 +37,4 @@ docker run -it \
| whisper | https://www.modelscope.cn/models/iic/Whisper-large-v3 | 23.8337 | 22.9085 | 0.910150 | 0.910150 | |
| paraformer | https://modelscope.cn/models/iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch | 4.7246 | 4.7719 | 0.955075 | 0.955075 | |
| conformer | https://www.modelscope.cn/models/iic/speech_conformer_asr_nat-zh-cn-16k-aishell2-vocab5212-pytorch | 95.9631 | 125.8649 | 0.349418 | 0.346090 | |
| uni_asr | https://www.modelscope.cn/models/iic/speech_UniASR-large_asr_2pass-zh-cn-16k-common-vocab8358-tensorflow1-offline | 70.5289 | 88.9481 | 0.717138 | 0.717138 | 该部分的适配修改了一些funASR源码 |
| uni_asr | https://www.modelscope.cn/models/iic/speech_UniASR-large_asr_2pass-zh-cn-16k-common-vocab8358-tensorflow1-offline | 70.5289 | 88.9481 | 0.717138 | 0.717138 | 该部分的适配修改了一些funASR源码 |

762
replaced_files/mr_v100/cif_predictor.py
View File

@@ -1,762 +0,0 @@
#!/usr/bin/env python3
# -*- encoding: utf-8 -*-
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
# MIT License (https://opensource.org/licenses/MIT)
import torch
import logging
import numpy as np
from funasr.register import tables
from funasr.train_utils.device_funcs import to_device
from funasr.models.transformer.utils.nets_utils import make_pad_mask
from torch.cuda.amp import autocast
@tables.register("predictor_classes", "CifPredictor")
class CifPredictor(torch.nn.Module):
def __init__(
self,
idim,
l_order,
r_order,
threshold=1.0,
dropout=0.1,
smooth_factor=1.0,
noise_threshold=0,
tail_threshold=0.45,
):
super().__init__()
self.pad = torch.nn.ConstantPad1d((l_order, r_order), 0)
self.cif_conv1d = torch.nn.Conv1d(idim, idim, l_order + r_order + 1, groups=idim)
self.cif_output = torch.nn.Linear(idim, 1)
self.dropout = torch.nn.Dropout(p=dropout)
self.threshold = threshold
self.smooth_factor = smooth_factor
self.noise_threshold = noise_threshold
self.tail_threshold = tail_threshold
def forward(
self,
hidden,
target_label=None,
mask=None,
ignore_id=-1,
mask_chunk_predictor=None,
target_label_length=None,
):
with autocast(False):
h = hidden
context = h.transpose(1, 2)
queries = self.pad(context)
memory = self.cif_conv1d(queries)
output = memory + context
output = self.dropout(output)
output = output.transpose(1, 2)
output = torch.relu(output)
output = self.cif_output(output)
alphas = torch.sigmoid(output)
alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)
if mask is not None:
mask = mask.transpose(-1, -2).float()
alphas = alphas * mask
if mask_chunk_predictor is not None:
alphas = alphas * mask_chunk_predictor
alphas = alphas.squeeze(-1)
mask = mask.squeeze(-1)
if target_label_length is not None:
target_length = target_label_length
elif target_label is not None:
target_length = (target_label != ignore_id).float().sum(-1)
else:
target_length = None
token_num = alphas.sum(-1)
if target_length is not None:
alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))
elif self.tail_threshold > 0.0:
hidden, alphas, token_num = self.tail_process_fn(
hidden, alphas, token_num, mask=mask
)
acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)
if target_length is None and self.tail_threshold > 0.0:
token_num_int = torch.max(token_num).type(torch.int32).item()
acoustic_embeds = acoustic_embeds[:, :token_num_int, :]
return acoustic_embeds, token_num, alphas, cif_peak
def tail_process_fn(self, hidden, alphas, token_num=None, mask=None):
b, t, d = hidden.size()
tail_threshold = self.tail_threshold
if mask is not None:
zeros_t = torch.zeros((b, 1), dtype=torch.float32, device=alphas.device)
ones_t = torch.ones_like(zeros_t)
mask_1 = torch.cat([mask, zeros_t], dim=1)
mask_2 = torch.cat([ones_t, mask], dim=1)
mask = mask_2 - mask_1
tail_threshold = mask * tail_threshold
alphas = torch.cat([alphas, zeros_t], dim=1)
alphas = torch.add(alphas, tail_threshold)
else:
tail_threshold = torch.tensor([tail_threshold], dtype=alphas.dtype).to(alphas.device)
tail_threshold = torch.reshape(tail_threshold, (1, 1))
alphas = torch.cat([alphas, tail_threshold], dim=1)
zeros = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device)
hidden = torch.cat([hidden, zeros], dim=1)
token_num = alphas.sum(dim=-1)
token_num_floor = torch.floor(token_num)
return hidden, alphas, token_num_floor
def gen_frame_alignments(
self, alphas: torch.Tensor = None, encoder_sequence_length: torch.Tensor = None
):
batch_size, maximum_length = alphas.size()
int_type = torch.int32
is_training = self.training
if is_training:
token_num = torch.round(torch.sum(alphas, dim=1)).type(int_type)
else:
token_num = torch.floor(torch.sum(alphas, dim=1)).type(int_type)
max_token_num = torch.max(token_num).item()
alphas_cumsum = torch.cumsum(alphas, dim=1)
alphas_cumsum = torch.floor(alphas_cumsum).type(int_type)
alphas_cumsum = alphas_cumsum[:, None, :].repeat(1, max_token_num, 1)
index = torch.ones([batch_size, max_token_num], dtype=int_type)
index = torch.cumsum(index, dim=1)
index = index[:, :, None].repeat(1, 1, maximum_length).to(alphas_cumsum.device)
index_div = torch.floor(torch.true_divide(alphas_cumsum, index)).type(int_type)
index_div_bool_zeros = index_div.eq(0)
index_div_bool_zeros_count = torch.sum(index_div_bool_zeros, dim=-1) + 1
index_div_bool_zeros_count = torch.clamp(
index_div_bool_zeros_count, 0, encoder_sequence_length.max()
)
token_num_mask = (~make_pad_mask(token_num, maxlen=max_token_num)).to(token_num.device)
index_div_bool_zeros_count *= token_num_mask
index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(
1, 1, maximum_length
)
ones = torch.ones_like(index_div_bool_zeros_count_tile)
zeros = torch.zeros_like(index_div_bool_zeros_count_tile)
ones = torch.cumsum(ones, dim=2)
cond = index_div_bool_zeros_count_tile == ones
index_div_bool_zeros_count_tile = torch.where(cond, zeros, ones)
index_div_bool_zeros_count_tile_bool = index_div_bool_zeros_count_tile.type(torch.bool)
index_div_bool_zeros_count_tile = 1 - index_div_bool_zeros_count_tile_bool.type(int_type)
index_div_bool_zeros_count_tile_out = torch.sum(index_div_bool_zeros_count_tile, dim=1)
index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out.type(int_type)
predictor_mask = (
(~make_pad_mask(encoder_sequence_length, maxlen=encoder_sequence_length.max()))
.type(int_type)
.to(encoder_sequence_length.device)
)
index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out * predictor_mask
predictor_alignments = index_div_bool_zeros_count_tile_out
predictor_alignments_length = predictor_alignments.sum(-1).type(
encoder_sequence_length.dtype
)
return predictor_alignments.detach(), predictor_alignments_length.detach()
@tables.register("predictor_classes", "CifPredictorV2")
class CifPredictorV2(torch.nn.Module):
def __init__(
self,
idim,
l_order,
r_order,
threshold=1.0,
dropout=0.1,
smooth_factor=1.0,
noise_threshold=0,
tail_threshold=0.0,
tf2torch_tensor_name_prefix_torch="predictor",
tf2torch_tensor_name_prefix_tf="seq2seq/cif",
tail_mask=True,
):
super().__init__()
self.pad = torch.nn.ConstantPad1d((l_order, r_order), 0)
self.cif_conv1d = torch.nn.Conv1d(idim, idim, l_order + r_order + 1)
self.cif_output = torch.nn.Linear(idim, 1)
self.dropout = torch.nn.Dropout(p=dropout)
self.threshold = threshold
self.smooth_factor = smooth_factor
self.noise_threshold = noise_threshold
self.tail_threshold = tail_threshold
self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch
self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf
self.tail_mask = tail_mask
def forward(
self,
hidden,
target_label=None,
mask=None,
ignore_id=-1,
mask_chunk_predictor=None,
target_label_length=None,
):
with autocast(False):
h = hidden
context = h.transpose(1, 2)
queries = self.pad(context)
output = torch.relu(self.cif_conv1d(queries))
output = output.transpose(1, 2)
output = self.cif_output(output)
alphas = torch.sigmoid(output)
alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)
if mask is not None:
mask = mask.transpose(-1, -2).float()
alphas = alphas * mask
if mask_chunk_predictor is not None:
alphas = alphas * mask_chunk_predictor
alphas = alphas.squeeze(-1)
mask = mask.squeeze(-1)
if target_label_length is not None:
target_length = target_label_length.squeeze(-1)
elif target_label is not None:
target_length = (target_label != ignore_id).float().sum(-1)
else:
target_length = None
token_num = alphas.sum(-1)
if target_length is not None:
alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))
elif self.tail_threshold > 0.0:
if self.tail_mask:
hidden, alphas, token_num = self.tail_process_fn(
hidden, alphas, token_num, mask=mask
)
else:
hidden, alphas, token_num = self.tail_process_fn(
hidden, alphas, token_num, mask=None
)
acoustic_embeds, cif_peak = cif_v1(hidden, alphas, self.threshold)
if target_length is None and self.tail_threshold > 0.0:
token_num_int = torch.max(token_num).type(torch.int32).item()
acoustic_embeds = acoustic_embeds[:, :token_num_int, :]
return acoustic_embeds, token_num, alphas, cif_peak
def forward_chunk(self, hidden, cache=None, **kwargs):
is_final = kwargs.get("is_final", False)
batch_size, len_time, hidden_size = hidden.shape
h = hidden
context = h.transpose(1, 2)
queries = self.pad(context)
output = torch.relu(self.cif_conv1d(queries))
output = output.transpose(1, 2)
output = self.cif_output(output)
alphas = torch.sigmoid(output)
alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)
alphas = alphas.squeeze(-1)
token_length = []
list_fires = []
list_frames = []
cache_alphas = []
cache_hiddens = []
if cache is not None and "chunk_size" in cache:
alphas[:, : cache["chunk_size"][0]] = 0.0
if not is_final:
alphas[:, sum(cache["chunk_size"][:2]) :] = 0.0
if cache is not None and "cif_alphas" in cache and "cif_hidden" in cache:
cache["cif_hidden"] = to_device(cache["cif_hidden"], device=hidden.device)
cache["cif_alphas"] = to_device(cache["cif_alphas"], device=alphas.device)
hidden = torch.cat((cache["cif_hidden"], hidden), dim=1)
alphas = torch.cat((cache["cif_alphas"], alphas), dim=1)
if cache is not None and is_final:
tail_hidden = torch.zeros((batch_size, 1, hidden_size), device=hidden.device)
tail_alphas = torch.tensor([[self.tail_threshold]], device=alphas.device)
tail_alphas = torch.tile(tail_alphas, (batch_size, 1))
hidden = torch.cat((hidden, tail_hidden), dim=1)
alphas = torch.cat((alphas, tail_alphas), dim=1)
len_time = alphas.shape[1]
for b in range(batch_size):
integrate = 0.0
frames = torch.zeros((hidden_size), device=hidden.device)
list_frame = []
list_fire = []
for t in range(len_time):
alpha = alphas[b][t]
if alpha + integrate < self.threshold:
integrate += alpha
list_fire.append(integrate)
frames += alpha * hidden[b][t]
else:
frames += (self.threshold - integrate) * hidden[b][t]
list_frame.append(frames)
integrate += alpha
list_fire.append(integrate)
integrate -= self.threshold
frames = integrate * hidden[b][t]
cache_alphas.append(integrate)
if integrate > 0.0:
cache_hiddens.append(frames / integrate)
else:
cache_hiddens.append(frames)
token_length.append(torch.tensor(len(list_frame), device=alphas.device))
list_fires.append(list_fire)
list_frames.append(list_frame)
cache["cif_alphas"] = torch.stack(cache_alphas, axis=0)
cache["cif_alphas"] = torch.unsqueeze(cache["cif_alphas"], axis=0)
cache["cif_hidden"] = torch.stack(cache_hiddens, axis=0)
cache["cif_hidden"] = torch.unsqueeze(cache["cif_hidden"], axis=0)
max_token_len = max(token_length)
if max_token_len == 0:
return hidden, torch.stack(token_length, 0), None, None
list_ls = []
for b in range(batch_size):
pad_frames = torch.zeros(
(max_token_len - token_length[b], hidden_size), device=alphas.device
)
if token_length[b] == 0:
list_ls.append(pad_frames)
else:
list_frames[b] = torch.stack(list_frames[b])
list_ls.append(torch.cat((list_frames[b], pad_frames), dim=0))
cache["cif_alphas"] = torch.stack(cache_alphas, axis=0)
cache["cif_alphas"] = torch.unsqueeze(cache["cif_alphas"], axis=0)
cache["cif_hidden"] = torch.stack(cache_hiddens, axis=0)
cache["cif_hidden"] = torch.unsqueeze(cache["cif_hidden"], axis=0)
return torch.stack(list_ls, 0), torch.stack(token_length, 0), None, None
def tail_process_fn(self, hidden, alphas, token_num=None, mask=None):
b, t, d = hidden.size()
tail_threshold = self.tail_threshold
if mask is not None:
zeros_t = torch.zeros((b, 1), dtype=torch.float32, device=alphas.device)
ones_t = torch.ones_like(zeros_t)
mask_1 = torch.cat([mask, zeros_t], dim=1)
mask_2 = torch.cat([ones_t, mask], dim=1)
mask = mask_2 - mask_1
tail_threshold = mask * tail_threshold
alphas = torch.cat([alphas, zeros_t], dim=1)
alphas = torch.add(alphas, tail_threshold)
else:
tail_threshold = torch.tensor([tail_threshold], dtype=alphas.dtype).to(alphas.device)
tail_threshold = torch.reshape(tail_threshold, (1, 1))
if b > 1:
alphas = torch.cat([alphas, tail_threshold.repeat(b, 1)], dim=1)
else:
alphas = torch.cat([alphas, tail_threshold], dim=1)
zeros = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device)
hidden = torch.cat([hidden, zeros], dim=1)
token_num = alphas.sum(dim=-1)
token_num_floor = torch.floor(token_num)
return hidden, alphas, token_num_floor
def gen_frame_alignments(
self, alphas: torch.Tensor = None, encoder_sequence_length: torch.Tensor = None
):
batch_size, maximum_length = alphas.size()
int_type = torch.int32
is_training = self.training
if is_training:
token_num = torch.round(torch.sum(alphas, dim=1)).type(int_type)
else:
token_num = torch.floor(torch.sum(alphas, dim=1)).type(int_type)
max_token_num = torch.max(token_num).item()
alphas_cumsum = torch.cumsum(alphas, dim=1)
alphas_cumsum = torch.floor(alphas_cumsum).type(int_type)
alphas_cumsum = alphas_cumsum[:, None, :].repeat(1, max_token_num, 1)
index = torch.ones([batch_size, max_token_num], dtype=int_type)
index = torch.cumsum(index, dim=1)
index = index[:, :, None].repeat(1, 1, maximum_length).to(alphas_cumsum.device)
index_div = torch.floor(torch.true_divide(alphas_cumsum, index)).type(int_type)
index_div_bool_zeros = index_div.eq(0)
index_div_bool_zeros_count = torch.sum(index_div_bool_zeros, dim=-1) + 1
index_div_bool_zeros_count = torch.clamp(
index_div_bool_zeros_count, 0, encoder_sequence_length.max()
)
token_num_mask = (~make_pad_mask(token_num, maxlen=max_token_num)).to(token_num.device)
index_div_bool_zeros_count *= token_num_mask
index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(
1, 1, maximum_length
)
ones = torch.ones_like(index_div_bool_zeros_count_tile)
zeros = torch.zeros_like(index_div_bool_zeros_count_tile)
ones = torch.cumsum(ones, dim=2)
cond = index_div_bool_zeros_count_tile == ones
index_div_bool_zeros_count_tile = torch.where(cond, zeros, ones)
index_div_bool_zeros_count_tile_bool = index_div_bool_zeros_count_tile.type(torch.bool)
index_div_bool_zeros_count_tile = 1 - index_div_bool_zeros_count_tile_bool.type(int_type)
index_div_bool_zeros_count_tile_out = torch.sum(index_div_bool_zeros_count_tile, dim=1)
index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out.type(int_type)
predictor_mask = (
(~make_pad_mask(encoder_sequence_length, maxlen=encoder_sequence_length.max()))
.type(int_type)
.to(encoder_sequence_length.device)
)
index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out * predictor_mask
predictor_alignments = index_div_bool_zeros_count_tile_out
predictor_alignments_length = predictor_alignments.sum(-1).type(
encoder_sequence_length.dtype
)
return predictor_alignments.detach(), predictor_alignments_length.detach()
@tables.register("predictor_classes", "CifPredictorV2Export")
class CifPredictorV2Export(torch.nn.Module):
def __init__(self, model, **kwargs):
super().__init__()
self.pad = model.pad
self.cif_conv1d = model.cif_conv1d
self.cif_output = model.cif_output
self.threshold = model.threshold
self.smooth_factor = model.smooth_factor
self.noise_threshold = model.noise_threshold
self.tail_threshold = model.tail_threshold
def forward(
self,
hidden: torch.Tensor,
mask: torch.Tensor,
):
alphas, token_num = self.forward_cnn(hidden, mask)
mask = mask.transpose(-1, -2).float()
mask = mask.squeeze(-1)
hidden, alphas, token_num = self.tail_process_fn(hidden, alphas, mask=mask)
acoustic_embeds, cif_peak = cif_v1_export(hidden, alphas, self.threshold)
return acoustic_embeds, token_num, alphas, cif_peak
def forward_cnn(
self,
hidden: torch.Tensor,
mask: torch.Tensor,
):
h = hidden
context = h.transpose(1, 2)
queries = self.pad(context)
output = torch.relu(self.cif_conv1d(queries))
output = output.transpose(1, 2)
output = self.cif_output(output)
alphas = torch.sigmoid(output)
alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)
mask = mask.transpose(-1, -2).float()
alphas = alphas * mask
alphas = alphas.squeeze(-1)
token_num = alphas.sum(-1)
return alphas, token_num
def tail_process_fn(self, hidden, alphas, token_num=None, mask=None):
b, t, d = hidden.size()
tail_threshold = self.tail_threshold
zeros_t = torch.zeros((b, 1), dtype=torch.float32, device=alphas.device)
ones_t = torch.ones_like(zeros_t)
mask_1 = torch.cat([mask, zeros_t], dim=1)
mask_2 = torch.cat([ones_t, mask], dim=1)
mask = mask_2 - mask_1
tail_threshold = mask * tail_threshold
alphas = torch.cat([alphas, zeros_t], dim=1)
alphas = torch.add(alphas, tail_threshold)
zeros = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device)
hidden = torch.cat([hidden, zeros], dim=1)
token_num = alphas.sum(dim=-1)
token_num_floor = torch.floor(token_num)
return hidden, alphas, token_num_floor
@torch.jit.script
def cif_v1_export(hidden, alphas, threshold: float):
device = hidden.device
dtype = hidden.dtype
batch_size, len_time, hidden_size = hidden.size()
threshold = torch.tensor([threshold], dtype=alphas.dtype).to(alphas.device)
frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)
fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)
# prefix_sum = torch.cumsum(alphas, dim=1)
prefix_sum = torch.cumsum(alphas, dim=1, dtype=torch.float64).to(
torch.float32
) # cumsum precision degradation cause wrong result in extreme
prefix_sum_floor = torch.floor(prefix_sum)
dislocation_prefix_sum = torch.roll(prefix_sum, 1, dims=1)
dislocation_prefix_sum_floor = torch.floor(dislocation_prefix_sum)
dislocation_prefix_sum_floor[:, 0] = 0
dislocation_diff = prefix_sum_floor - dislocation_prefix_sum_floor
fire_idxs = dislocation_diff > 0
fires[fire_idxs] = 1
fires = fires + prefix_sum - prefix_sum_floor
# prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)
prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).repeat((1, 1, hidden_size)) * hidden, dim=1)
frames = prefix_sum_hidden[fire_idxs]
shift_frames = torch.roll(frames, 1, dims=0)
batch_len = fire_idxs.sum(1)
batch_idxs = torch.cumsum(batch_len, dim=0)
shift_batch_idxs = torch.roll(batch_idxs, 1, dims=0)
shift_batch_idxs[0] = 0
shift_frames[shift_batch_idxs] = 0
remains = fires - torch.floor(fires)
# remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
remain_frames = remains[fire_idxs].unsqueeze(-1).repeat((1, hidden_size)) * hidden[fire_idxs]
shift_remain_frames = torch.roll(remain_frames, 1, dims=0)
shift_remain_frames[shift_batch_idxs] = 0
frames = frames - shift_frames + shift_remain_frames - remain_frames
# max_label_len = batch_len.max()
max_label_len = alphas.sum(dim=-1)
max_label_len = torch.floor(max_label_len).max().to(dtype=torch.int64)
# frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)
frame_fires_idxs = indices < batch_len.unsqueeze(1)
frame_fires[frame_fires_idxs] = frames
return frame_fires, fires
@torch.jit.script
def cif_export(hidden, alphas, threshold: float):
batch_size, len_time, hidden_size = hidden.size()
threshold = torch.tensor([threshold], dtype=alphas.dtype).to(alphas.device)
# loop varss
integrate = torch.zeros([batch_size], dtype=alphas.dtype, device=hidden.device)
frame = torch.zeros([batch_size, hidden_size], dtype=hidden.dtype, device=hidden.device)
# intermediate vars along time
list_fires = []
list_frames = []
for t in range(len_time):
alpha = alphas[:, t]
distribution_completion = (
torch.ones([batch_size], dtype=alphas.dtype, device=hidden.device) - integrate
)
integrate += alpha
list_fires.append(integrate)
fire_place = integrate >= threshold
integrate = torch.where(
fire_place,
integrate - torch.ones([batch_size], dtype=alphas.dtype, device=hidden.device),
integrate,
)
cur = torch.where(fire_place, distribution_completion, alpha)
remainds = alpha - cur
frame += cur[:, None] * hidden[:, t, :]
list_frames.append(frame)
frame = torch.where(
fire_place[:, None].repeat(1, hidden_size), remainds[:, None] * hidden[:, t, :], frame
)
fires = torch.stack(list_fires, 1)
frames = torch.stack(list_frames, 1)
fire_idxs = fires >= threshold
frame_fires = torch.zeros_like(hidden)
max_label_len = frames[0, fire_idxs[0]].size(0)
for b in range(batch_size):
frame_fire = frames[b, fire_idxs[b]]
frame_len = frame_fire.size(0)
frame_fires[b, :frame_len, :] = frame_fire
if frame_len >= max_label_len:
max_label_len = frame_len
frame_fires = frame_fires[:, :max_label_len, :]
return frame_fires, fires
class mae_loss(torch.nn.Module):
def __init__(self, normalize_length=False):
super(mae_loss, self).__init__()
self.normalize_length = normalize_length
self.criterion = torch.nn.L1Loss(reduction="sum")
def forward(self, token_length, pre_token_length):
loss_token_normalizer = token_length.size(0)
if self.normalize_length:
loss_token_normalizer = token_length.sum().type(torch.float32)
loss = self.criterion(token_length, pre_token_length)
loss = loss / loss_token_normalizer
return loss
def cif(hidden, alphas, threshold):
batch_size, len_time, hidden_size = hidden.size()
# loop varss
integrate = torch.zeros([batch_size], device=hidden.device)
frame = torch.zeros([batch_size, hidden_size], device=hidden.device)
# intermediate vars along time
list_fires = []
list_frames = []
for t in range(len_time):
alpha = alphas[:, t]
distribution_completion = torch.ones([batch_size], device=hidden.device) - integrate
integrate += alpha
list_fires.append(integrate)
fire_place = integrate >= threshold
integrate = torch.where(
fire_place, integrate - torch.ones([batch_size], device=hidden.device), integrate
)
cur = torch.where(fire_place, distribution_completion, alpha)
remainds = alpha - cur
frame += cur[:, None] * hidden[:, t, :]
list_frames.append(frame)
frame = torch.where(
fire_place[:, None].repeat(1, hidden_size), remainds[:, None] * hidden[:, t, :], frame
)
fires = torch.stack(list_fires, 1)
frames = torch.stack(list_frames, 1)
list_ls = []
len_labels = torch.round(alphas.sum(-1)).int()
max_label_len = len_labels.max()
for b in range(batch_size):
fire = fires[b, :]
l = torch.index_select(frames[b, :, :], 0, torch.nonzero(fire >= threshold).squeeze())
pad_l = torch.zeros([max_label_len - l.size(0), hidden_size], device=hidden.device)
list_ls.append(torch.cat([l, pad_l], 0))
return torch.stack(list_ls, 0), fires
def cif_wo_hidden_v1(alphas, threshold, return_fire_idxs=False):
batch_size, len_time = alphas.size()
device = alphas.device
dtype = alphas.dtype
threshold = torch.tensor([threshold], dtype=alphas.dtype).to(alphas.device)
fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)
if torch.cuda.get_device_name() == "Iluvatar MR-V100":
prefix_sum = torch.cumsum(alphas, dim=1)
else:
prefix_sum = torch.cumsum(alphas, dim=1, dtype=torch.float64).to(
torch.float32
) # cumsum precision degradation cause wrong result in extreme
prefix_sum_floor = torch.floor(prefix_sum)
dislocation_prefix_sum = torch.roll(prefix_sum, 1, dims=1)
dislocation_prefix_sum_floor = torch.floor(dislocation_prefix_sum)
dislocation_prefix_sum_floor[:, 0] = 0
dislocation_diff = prefix_sum_floor - dislocation_prefix_sum_floor
fire_idxs = dislocation_diff > 0
fires[fire_idxs] = 1
fires = fires + prefix_sum - prefix_sum_floor
if return_fire_idxs:
return fires, fire_idxs
return fires
def cif_v1(hidden, alphas, threshold):
fires, fire_idxs = cif_wo_hidden_v1(alphas, threshold, return_fire_idxs=True)
device = hidden.device
dtype = hidden.dtype
batch_size, len_time, hidden_size = hidden.size()
# frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)
# prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)
frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)
prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).repeat((1, 1, hidden_size)) * hidden, dim=1)
frames = prefix_sum_hidden[fire_idxs]
shift_frames = torch.roll(frames, 1, dims=0)
batch_len = fire_idxs.sum(1)
batch_idxs = torch.cumsum(batch_len, dim=0)
shift_batch_idxs = torch.roll(batch_idxs, 1, dims=0)
shift_batch_idxs[0] = 0
shift_frames[shift_batch_idxs] = 0
remains = fires - torch.floor(fires)
# remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
remain_frames = remains[fire_idxs].unsqueeze(-1).repeat((1, hidden_size)) * hidden[fire_idxs]
shift_remain_frames = torch.roll(remain_frames, 1, dims=0)
shift_remain_frames[shift_batch_idxs] = 0
frames = frames - shift_frames + shift_remain_frames - remain_frames
# max_label_len = batch_len.max()
max_label_len = (
torch.round(alphas.sum(-1)).int().max()
) # torch.round to calculate the max length
# frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)
frame_fires_idxs = indices < batch_len.unsqueeze(1)
frame_fires[frame_fires_idxs] = frames
return frame_fires, fires
def cif_wo_hidden(alphas, threshold):
batch_size, len_time = alphas.size()
# loop varss
integrate = torch.zeros([batch_size], device=alphas.device)
# intermediate vars along time
list_fires = []
for t in range(len_time):
alpha = alphas[:, t]
integrate += alpha
list_fires.append(integrate)
fire_place = integrate >= threshold
integrate = torch.where(
fire_place,
integrate - torch.ones([batch_size], device=alphas.device) * threshold,
integrate,
)
fires = torch.stack(list_fires, 1)
return fires

28
test_funasr.py
View File

@@ -1,7 +1,7 @@
import os
import time
import torchaudio
import torch
import torchaudio
from funasr import AutoModel
from funasr.utils.postprocess_utils import rich_transcription_postprocess
from utils.calculate import cal_per_cer
@@ -34,18 +34,11 @@ def test_funasr(model_dir, audio_file, answer_file, use_gpu):
model_name = os.path.basename(model_dir)
model_type = determine_model_type(model_name)
if torch.cuda.get_device_name() == "Iluvatar BI-V100" and model_type == "whisper":
# 天垓100情况下的Whisper需要绕过不支持算子
torch.backends.cuda.enable_flash_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp(False)
torch.backends.cuda.enable_math_sdp(True)
# 不使用VAD, punct，spk模型，就测试原始ASR能力
model = AutoModel(
model=model_dir,
# vad_model="fsmn-vad",
# vad_kwargs={"max_single_segment_time": 30000},
vad_model=None,
device="cuda:0" if use_gpu else "cpu",
disable_update=True
)
@@ -114,14 +107,6 @@ def test_funasr(model_dir, audio_file, answer_file, use_gpu):
batch_size_s=300
)
text = res[0]["text"]
# elif model_type == "uni_asr":
# if i == 0:
# os.remove(segment_path)
# continue
# res = model.generate(
# input=segment_path
# )
# text = res[0]["text"]
else:
raise RuntimeError("unknown model type")
ts2 = time.time()
@@ -142,6 +127,13 @@ def test_funasr(model_dir, audio_file, answer_file, use_gpu):
return processing_time, acc, generated_text
if __name__ == "__main__":
if torch.cuda.is_available():
cuda_tensor = torch.randn(2, 2, device='cuda:0')
print(f"CUDA device index: {cuda_tensor.get_device()}")
else:
print("CUDA not available")
os._exit(1)
test_result = {
"time_cuda": 0,
"acc_cuda": 0,
@@ -176,5 +168,5 @@ if __name__ == "__main__":
json.dump(test_result, fp, ensure_ascii=False, indent=4)
# 如果是SUT起来镜像的话，需要加上下面让pod永不停止以迎合k8s deployment, 本地测试以及docker run均不需要
if K8S_TEST:
print(f"Start to sleep indefinitely", flush=True)
time.sleep(100000)
print(f"Start to sleep indefinity", flush=True)
time.sleep(100000)