3D-Speaker模型微调

Luke Ewin

于 2025-04-12 05:32:36 发布

阅读量322

点赞数 9

本文链接：https://blog.csdn.net/qq_43907505/article/details/147157706

版权

声纹识别专栏收录该内容

1 篇文章

订阅专栏

0. 下载数据

使用多线程下载器下载

apt-get update -y
apt-get install aria2 -y

开始下载数据（-s 16表示开启16个线程，-x 16表示每个url分配16个线程）

下载到/root/autodl-tmp/3D-Speaker/egs/cnceleb/sv-cam++/data/download_data目录中

aria2c -s 16 -x 16 -c https://openslr.elda.org/resources/17/musan.tar.gz
aria2c -s 16 -x 16 -c https://us.openslr.org/resources/28/rirs_noises.zip
aria2c -s 16 -x 16 -c https://www.openslr.org/resources/82/cn-celeb_v2.tar.gz
aria2c -s 16 -x 16 -c https://www.openslr.org/resources/82/cn-celeb2_v2.tar.gzaa
aria2c -s 16 -x 16 -c https://www.openslr.org/resources/82/cn-celeb2_v2.tar.gzab
aria2c -s 16 -x 16 -c https://www.openslr.org/resources/82/cn-celeb2_v2.tar.gzac

注意后面三个是分块下载的，下载完成之后需要合并在一起。

cat cn-celeb2_v2.tar.gza* >cn-celeb2_v2.tar.gz

如果你使用的python版本是3.12，那么按照依赖的时候会报下面错误，这是因为python版本过高的原因，使用python>=3.8 && python<=3.11。

root@autodl-container-b3e54da89e-80ab5eb3:~/autodl-tmp/3D-Speaker# pip install -r requirements.txt 
Looking in indexes: http://mirrors.aliyun.com/pypi/simple
Requirement already satisfied: tqdm>=4.42.0 in /root/miniconda3/lib/python3.12/site-packages (from -r requirements.txt (line 1)) (4.66.2)
Collecting scipy>=1.7.0 (from -r requirements.txt (line 2))
  Downloading http://mirrors.aliyun.com/pypi/packages/c0/53/eaada1a414c026673eb983f8b4a55fe5eb172725d33d62c1b21f63ff6ca4/scipy-1.15.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (37.3 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 37.3/37.3 MB 10.3 MB/s eta 0:00:00
Collecting numpy<1.24,>=1.20.0 (from -r requirements.txt (line 3))
  Using cached http://mirrors.aliyun.com/pypi/packages/42/38/775b43da55fa7473015eddc9a819571517d9a271a9f8134f68fb9be2f212/numpy-1.23.5.tar.gz (10.7 MB)
  Installing build dependencies ... done
  Getting requirements to build wheel ... error
  error: subprocess-exited-with-error
  
  × Getting requirements to build wheel did not run successfully.
  │ exit code: 1
  ╰─> [33 lines of output]
      Traceback (most recent call last):
        File "/root/miniconda3/lib/python3.12/site-packages/pip/_vendor/pyproject_hooks/_in_process/_in_process.py", line 353, in <module>
          main()
        File "/root/miniconda3/lib/python3.12/site-packages/pip/_vendor/pyproject_hooks/_in_process/_in_process.py", line 335, in main
          json_out['return_val'] = hook(**hook_input['kwargs'])
                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        File "/root/miniconda3/lib/python3.12/site-packages/pip/_vendor/pyproject_hooks/_in_process/_in_process.py", line 112, in get_requires_for_build_wheel
          backend = _build_backend()
                    ^^^^^^^^^^^^^^^^
        File "/root/miniconda3/lib/python3.12/site-packages/pip/_vendor/pyproject_hooks/_in_process/_in_process.py", line 77, in _build_backend
          obj = import_module(mod_path)
                ^^^^^^^^^^^^^^^^^^^^^^^
        File "/root/miniconda3/lib/python3.12/importlib/__init__.py", line 90, in import_module
          return _bootstrap._gcd_import(name[level:], package, level)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        File "<frozen importlib._bootstrap>", line 1387, in _gcd_import
        File "<frozen importlib._bootstrap>", line 1360, in _find_and_load
        File "<frozen importlib._bootstrap>", line 1310, in _find_and_load_unlocked
        File "<frozen importlib._bootstrap>", line 488, in _call_with_frames_removed
        File "<frozen importlib._bootstrap>", line 1387, in _gcd_import
        File "<frozen importlib._bootstrap>", line 1360, in _find_and_load
        File "<frozen importlib._bootstrap>", line 1331, in _find_and_load_unlocked
        File "<frozen importlib._bootstrap>", line 935, in _load_unlocked
        File "<frozen importlib._bootstrap_external>", line 995, in exec_module
        File "<frozen importlib._bootstrap>", line 488, in _call_with_frames_removed
        File "/tmp/pip-build-env-vkyrguzl/overlay/lib/python3.12/site-packages/setuptools/__init__.py", line 16, in <module>
          import setuptools.version
        File "/tmp/pip-build-env-vkyrguzl/overlay/lib/python3.12/site-packages/setuptools/version.py", line 1, in <module>
          import pkg_resources
        File "/tmp/pip-build-env-vkyrguzl/overlay/lib/python3.12/site-packages/pkg_resources/__init__.py", line 2172, in <module>
          register_finder(pkgutil.ImpImporter, find_on_path)
                          ^^^^^^^^^^^^^^^^^^^
      AttributeError: module 'pkgutil' has no attribute 'ImpImporter'. Did you mean: 'zipimporter'?
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
error: subprocess-exited-with-error

× Getting requirements to build wheel did not run successfully.
│ exit code: 1
╰─> See above for output.

note: This error originates from a subprocess, and is likely not a problem with pip.

1. 创建环境

git clone https://github.com/modelscope/3D-Speaker.git && cd 3D-Speaker
conda create -n 3D-Speaker python=3.8
conda activate 3D-Speaker
pip install -r requirements.txt

注意这里我把requirements.txt中的torch和torchaudio删了，需要执行下面命令安装。

pip3 install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 -f  https://mirrors.aliyun.com/pytorch-wheels/cu118

pip install scipy

2. 修改配置

根据自己租的显卡数量来修改run.sh脚本，并且修改conf/cam++.yaml配置文件。

下面是run.sh脚本。

#!/bin/bash
# Copyright 3D-Speaker (https://github.com/alibaba-damo-academy/3D-Speaker). All Rights Reserved.
# Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

set -e
. ./path.sh || exit 1

stage=1
stop_stage=5

data=data
exp=exp
exp_name=cam++
gpus="0 1 2 3 4 5"

. utils/parse_options.sh || exit 1

exp_dir=$exp/$exp_name

if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
  # In this stage we prepare the raw datasets, including CNCeleb1 and CNCeleb2.
  echo "Stage1: Preparing CN-Celeb dataset..."
  ./local/prepare_data_cncb.sh --stage 1 --stop_stage 4 --data ${data}
fi

if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
  # In this stage we prepare the data index files for training.
  echo "Stage2: Preparing training data index files..."
  python local/prepare_data_csv.py --data_dir $data/cnceleb_train
fi

if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
  # Train the speaker embedding model.
  echo "Stage3: Training the speaker model..."
  num_gpu=$(echo $gpus | awk -F ' ' '{print NF}')
  torchrun --nproc_per_node=$num_gpu --master_port=29501 speakerlab/bin/train.py --config conf/cam++.yaml --gpu $gpus \
           --data $data/cnceleb_train/train.csv --noise $data/musan/wav.scp --reverb $data/rirs/wav.scp --exp_dir $exp_dir
fi


if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
  # Extract embeddings of test datasets.
  echo "Stage4: Extracting speaker embeddings..."
  nj=12
  torchrun --nproc_per_node=$nj --master_port=29501 speakerlab/bin/extract.py --exp_dir $exp_dir \
           --data $data/cnceleb_test/wav.scp --use_gpu --gpu $gpus
fi

if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
  # Output score metrics.
  echo "Stage5: Computing score metrics..."
  trials="$data/cnceleb_test/trials"
  python speakerlab/bin/compute_score_metrics.py --enrol_data $exp_dir/embeddings --test_data $exp_dir/embeddings \
                                                 --scores_dir $exp_dir/scores --trials $trials
fi

下面是conf/cam++.yaml配置文件。

# Training config

# inputs
data:
noise:
reverb:

# outputs
exp_dir:

# basic
num_epoch: 75
save_epoch_freq: 5
log_batch_freq: 100

wav_len: 3.0 # duration(s) for each training sample.
sample_rate: 16000
aug_prob: 0.8
speed_pertub: True
lr: 0.2
min_lr: 0.00005

# dataloader
batch_size: 1024
num_workers: 16

# model
fbank_dim: 80
embedding_size: 512
num_classes: 2793


wav_reader:
  obj: speakerlab.process.processor.WavReader
  args:
    duration: <wav_len>
    sample_rate: <sample_rate>
    speed_pertub: <speed_pertub>

label_encoder:
  obj: speakerlab.process.processor.SpkLabelEncoder
  args:
    data_file: <data>

feature_extractor:
  obj: speakerlab.process.processor.FBank
  args:
    n_mels: <fbank_dim>
    sample_rate: <sample_rate>
    mean_nor: True

augmentations:
  obj: speakerlab.process.processor.SpkVeriAug
  args:
    aug_prob: <aug_prob>
    noise_file: <noise>
    reverb_file: <reverb>

preprocessor:
  wav_reader: <wav_reader>
  label_encoder: <label_encoder>
  augmentations: <augmentations>
  feature_extractor: <feature_extractor>

epoch_counter:
  obj: speakerlab.utils.epoch.EpochCounter
  args:
    limit: <num_epoch>

dataset:
  obj: speakerlab.dataset.dataset.WavSVDataset
  args:
    data_file: <data>
    preprocessor: <preprocessor>

dataloader:
  obj: torch.utils.data.DataLoader
  args:
    dataset: <dataset>
    batch_size: <batch_size>
    num_workers: <num_workers>
    pin_memory: True
    drop_last: True

embedding_model:
  obj: speakerlab.models.campplus.DTDNN.CAMPPlus
  args:
    feat_dim: <fbank_dim>
    embedding_size: <embedding_size>

classifier:
  obj: speakerlab.models.campplus.classifier.CosineClassifier
  args:
    input_dim: <embedding_size>
    out_neurons: <num_classes>

optimizer:
  obj: torch.optim.SGD
  args:
    params:
    lr: <lr>
    momentum: 0.9
    nesterov: True
    weight_decay: 0.0001

lr_scheduler:
  obj: speakerlab.process.scheduler.WarmupCosineScheduler
  args:
    optimizer: <optimizer>
    min_lr: <min_lr>
    max_lr: <lr>
    warmup_epoch: 5
    fix_epoch: <num_epoch>
    step_per_epoch:

loss:
  obj: speakerlab.loss.margin_loss.ArcMarginLoss
  args:
    scale: 32.0
    margin: 0.3
    easy_margin: False

margin_scheduler:
  obj: speakerlab.process.scheduler.MarginScheduler
  args:
    criterion: <loss>
    initial_margin: 0.0
    final_margin: 0.3
    increase_start_epoch: 20
    fix_epoch: 50
    step_per_epoch:

checkpointer:
  obj: speakerlab.utils.checkpoint.Checkpointer
  args:
    checkpoints_dir: <exp_dir>/models
    recoverables:
      embedding_model: <embedding_model>
      classifier: <classifier>
      epoch_counter: <epoch_counter>