Espent环境配置与实践

最新推荐文章于 2024-08-15 16:33:35 发布
最新推荐文章于 2024-08-15 16:33:35 发布
阅读量1k 收藏 16
点赞数 24
分类专栏： 语音识别 文章标签： 语音识别 pytorch
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/weixin_46560570/article/details/141107205
版权

Espent环境配置与实践

代码来源：espnet/espnet: End-to-End Speech Processing Toolkit (github.com)

官网文档：Common usages — ESPnet 202402 documentation

1. 环境配置

通过官方文档可得我们指定 PyTorch=1.12.1 CUDA=11.6。我们还支持许多其他版本。请查看 https://github.com/espnet/espnet/blob/master/tools/installers/install_torch.sh 以获取详细的版本列表。

安装CUDA=11.6

# 下载
# cuda的安装和配置：https://blog.csdn.net/weixin_46560570/article/details/140754242?spm=1001.2014.3001.5501
wget https://developer.download.nvidia.com/compute/cuda/11.6.2/local_installers/cuda_11.6.2_510.47.03_linux.run
# 安装
sh ./cuda_12.2.1_535.86.10_linux.run \
  --silent \
  --toolkit \
  --installpath=/s6home/lnj524/module/cuda/cuda-11.6 \
  --no-opengl-libs \
  --no-drm \
  --no-man-page
# 添加环境变量
vim ~/.bashrc
#将下方内容写入.bashrc 
export CUDA_HOME=/s6home/lnj524/module/cuda/cuda-11.6
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=$CUDA_HOME/extras/CUPTI/lib64:$LD_LIBRARY_PATH
export CUDAToolkit_ROOT_DIR=$CUDA_HOME
export CUDAToolkit_ROOT=$CUDA_HOME

export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
export CUDA_TOOLKIT_ROOT=$CUDA_HOME
export CUDA_BIN_PATH=$CUDA_HOME
export CUDA_PATH=$CUDA_HOME
export CUDA_INC_PATH=$CUDA_HOME/targets/x86_64-linux
export CFLAGS=-I$CUDA_HOME/targets/x86_64-linux/include:$CFLAGS
export CUDAToolkit_TARGET_DIR=$CUDA_HOME/targets/x86_64-linux

# 更新用户环境
source ~/.bashrc
# 验证
nvcc -V

安装Pytorch

# 克隆espnet代码
git clone https://gitee.com/chengsili/espnet.git
# 创建虚拟环境
# conda的安装和配置：https://blog.csdn.net/weixin_46560570/article/details/140754242?spm=1001.2014.3001.5501
conda create -n espnet python=3.9
# 激活虚拟环境
conda activate espnet
# 安装Pytorch
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
# 安装cudnn
conda install cudnn

项目配置

cd /s6home/lnj524/module/espnet/tools
# 链接cuda
./setup_cuda_env.sh /home/lnj524/module/cuda/cuda-11.6
# 链接python
./setup_python.sh $(command -v python3)
# 此步骤后，tools下生成activate_python.sh

#查看activate_python.sh
cat activate_python.sh

#!/usr/bin/env bash
# THIS FILE IS GENERATED BY tools/setup_python.sh
export PYTHONUSERBASE="/home/lnj524/module/espnet/tools/python_user_base"
export PATH="/home/lnj524/module/espnet/tools/python_user_base/bin":${PATH}
export PATH=/home/lnj524/miniconda3/envs/espnet/bin:${PATH}
export NLTK_DATA="${PYTHONUSERBASE}/nltk_data" # NLTK_DATA环境变量是后添上的，前三个变量是自动生成的。
# 报错记录
# NLTK_DATA环境变量，如果没有这个，nltk_data文件会产生在用户文件夹下。
# 若报错XXX.zip错误，自行安装和解压。
# https://gitee.com/qwererer2/nltk_data/blob/gh-pages/packages/corpora/cmudict.zip
# https://gitee.com/qwererer2/nltk_data/blob/gh-pages/packages/taggers/averaged_perceptron_tagger.zip

安装相关依赖

# 踩坑记录
# 服务器无法访问github，使用make安装时，网络连接超时。
# 解决方法:
cd /home/lnj524/module/espnet/tools/installers
# 检查文件目录下所有的.sh文件，将文件中的github上的项目url改成gitee上的项目url。
# 我改好的：https://gitee.com/chengsili/espent_tools_installers.git
# sph2pipe.sh中下载2.5.tar.gz使用：https://gitee.com/chengsili/sph2pipe/repository/archive/2.5.tar.gz
# SCTK.sh中下载9688a26.tar.gz使用：https://gitee.com/chengsili/SCTK/repository/archive/9688a26.tar.gz
# kenlm.sh中boost服务器可能下载不下来，如果你遇到了请看下文运行aishell，报错kenlm找不到。

cd /s6home/lnj524/module/espnet/tools
make TH_VERSION=1.12.1 CUDA_VERSION=11.6

pip install cmake
pip install sox

# 到此环境配置完成
# 检查安装，具体细节请访问：https://espnet.github.io/espnet/tutorial.html#transducer
python3 check_install.py

2. an4数据集

按照官网进行测试即可，注意Stage 11: ASR Training，设置GPU数量和使用的GPU编号

# 使用GPU 0和GPU 1
export CUDA_VISIBLE_DEVICES=0,1
# 训练
./asr.sh --stage 11 --stop_stage 11 --train_set train_nodev --valid_set train_dev --test_sets "train_dev test" --ngpu 2 --asr_config conf/train_asr_demo_transformer.yaml

3.aishell数据集

3.1kenlm报错
# 踩坑记录 - 安装kenlm
# 使用pip或者conda进行安装kenlm。使用conda list可看到kenlm=0.2.0，但是运行 ./asr.sh --stage 1 --stop_stage 1 .... 
# 仍然报错找不到kenlm。
# 解决方法，使用espnet提供的脚本。进入installers文件夹，运行./install_kenlm.sh报错，网络连接超时，可自行下载。
# https://boostorg.jfrog.io/artifactory/main/release/1.81.0/source/boost_1_81_0.tar.bz2
# 然后自行解压：tar xvf boost_1_81_0.tar.bz2，随后更改install_kenlm.sh，再次运行。 
# 具体如下：
cd /home/lnj524/module/espnet/tools/installers

# 将你下载好的压缩包放在installers目录下

tar xvf boost_1_81_0.tar.bz2

vim install_kenlm.sh

#更改后的install_kenlm.sh

#!/usr/bin/env bash
set -euo pipefail

if [ $# != 0 ]; then
    echo "Usage: $0"
    exit 1;
fi

boost_version=1.81.0

if [ ! -d boost_${boost_version//./_}_build ]; then
    (
        set -euo pipefail
        cd boost_${boost_version//./_}
        ./bootstrap.sh
        ./b2 install --prefix=$(pwd)/../boost_${boost_version//./_}_build install
    )
fi

if [ ! -d kenlm ]; then
    git clone https://gitee.com/chengsili/kenlm.git
fi

(
    set -euo pipefail
    cd kenlm

    mkdir -p build
    (
        set -euo pipefail
        cd build && cmake -DCMAKE_PREFIX_PATH=$(pwd)/../../boost_${boost_version//./_}_build .. && make
    )
    (
        set -euo pipefail
        python3 -m pip install -e .
    )
)

# 安装完成后，发现kenlm文件在installers目录下，将其移动到tools下即可
cd /home/lnj524/module/espnet/tools/installers
mv kenlm/ ../
3.2分阶段进行
数据准备
# 按照脚本执行第一步，数据集下载时间较长，自行下载数据集，更改data.sh中的数据集路径。
vim /home/lnj524/module/espnet/egs2/aishell/asr1/local/data.sh
AISHELL=/home/lnj524/module/data/opensource_data/aishell
第 1 阶段：数据准备
# 第 1 阶段：数据准备
./asr.sh --stage 1 --stop_stage 1 --train_set train --valid_set dev --test_sets "dev test"
espnet/egs2/aishell/asr1/data$ ls
dev  local  test  token_list  train
espnet/egs2/aishell/asr1/data$ ll train
spk2utt # 每个说话者的 ID 及其所有的发言列表
text    # 每个发言的转录文本
utt2spk # 每个发言 ID 及其对应的说话者 ID
wav.scp # 列出音频文件的路径
第 3 阶段：格式化
# 第 3 阶段：格式化 wav.scp： data/ -> dump/raw
./asr.sh --stage 3 --stop_stage 3 --train_set train --valid_set dev --test_sets "dev test"

2024-08-09T18:40:02 (asr.sh:283:main) ./asr.sh --stage 3 --stop_stage 3 --train_set train --valid_set dev --test_sets dev test
2024-08-09T18:40:02 (asr.sh:321:main) Info: The valid_set 'dev' is included in the test_sets. '--eval_valid_set true' is set and 'dev' is removed from the test_sets
2024-08-09T18:40:02 (asr.sh:564:main) Skipped stages:  9 14 15 
2024-08-09T18:40:02 (asr.sh:614:main) Stage 3: Format wav.scp: data/ -> dump/raw
utils/copy_data_dir.sh: copied data from data/train to dump/raw/org/train
utils/validate_data_dir.sh: Successfully validated data-directory dump/raw/org/train
2024-08-09T18:40:04 (format_wav_scp.sh:46:main) scripts/audio/format_wav_scp.sh --nj 32 --cmd run.pl --audio-format flac --fs 16k --multi-columns-input false --multi-columns-output false data/train/wav.scp dump/raw/org/train
2024-08-09T18:40:05 (format_wav_scp.sh:118:main) [info]: without segments
2024-08-09T18:47:10 (format_wav_scp.sh:153:main) Successfully finished. [elapsed=426s]
utils/copy_data_dir.sh: copied data from data/dev to dump/raw/org/dev
utils/validate_data_dir.sh: Successfully validated data-directory dump/raw/org/dev
2024-08-09T18:47:10 (format_wav_scp.sh:46:main) scripts/audio/format_wav_scp.sh --nj 32 --cmd run.pl --audio-format flac --fs 16k --multi-columns-input false --multi-columns-output false data/dev/wav.scp dump/raw/org/dev
2024-08-09T18:47:10 (format_wav_scp.sh:118:main) [info]: without segments
2024-08-09T18:47:56 (format_wav_scp.sh:153:main) Successfully finished. [elapsed=46s]
utils/copy_data_dir.sh: copied data from data/test to dump/raw/test
utils/validate_data_dir.sh: Successfully validated data-directory dump/raw/test
2024-08-09T18:47:56 (format_wav_scp.sh:46:main) scripts/audio/format_wav_scp.sh --nj 32 --cmd run.pl --audio-format flac --fs 16k --multi-columns-input false --multi-columns-output false data/test/wav.scp dump/raw/test
2024-08-09T18:47:57 (format_wav_scp.sh:118:main) [info]: without segments
2024-08-09T18:48:10 (format_wav_scp.sh:153:main) Successfully finished. [elapsed=14s]
2024-08-09T18:48:10 (asr.sh:1809:main) Successfully finished. [elapsed=488s]
第 4 阶段：删除长/短数据
# 第 4 阶段：删除长/短数据：dump/raw/org -> dump/raw
./asr.sh --stage 4 --stop_stage 4 --train_set train --valid_set dev --test_sets "dev test"

2024-08-09T18:50:44 (asr.sh:283:main) ./asr.sh --stage 4 --stop_stage 4 --train_set train --valid_set dev --test_sets dev test
2024-08-09T18:50:44 (asr.sh:321:main) Info: The valid_set 'dev' is included in the test_sets. '--eval_valid_set true' is set and 'dev' is removed from the test_sets
2024-08-09T18:50:44 (asr.sh:564:main) Skipped stages:  9 14 15 
2024-08-09T18:50:44 (asr.sh:799:main) Stage 4: Remove long/short data: dump/raw/org -> dump/raw
utils/copy_data_dir.sh: copied data from dump/raw/org/train to dump/raw/train
utils/validate_data_dir.sh: Successfully validated data-directory dump/raw/train
fix_data_dir.sh: kept all 120098 utterances.
fix_data_dir.sh: old files are kept in dump/raw/train/.backup
utils/copy_data_dir.sh: copied data from dump/raw/org/dev to dump/raw/dev
utils/validate_data_dir.sh: Successfully validated data-directory dump/raw/dev
fix_data_dir.sh: kept all 14326 utterances.
fix_data_dir.sh: old files are kept in dump/raw/dev/.backup
2024-08-09T18:50:51 (asr.sh:1809:main) Successfully finished. [elapsed=7s]
第 5 阶段：生成token_list
# 第 5 阶段：使用 BPE 从 dump/raw/train/text 生成token_list。
# 此处需要在asr.sh设置nbpe=4234  
./asr.sh --stage 5 --stop_stage 5 --train_set train --valid_set dev --test_sets "dev test"

2024-08-09T18:52:48 (asr.sh:283:main) ./asr.sh --stage 5 --stop_stage 5 --train_set train --valid_set dev --test_sets dev test
2024-08-09T18:52:48 (asr.sh:321:main) Info: The valid_set 'dev' is included in the test_sets. '--eval_valid_set true' is set and 'dev' is removed from the test_sets
2024-08-09T18:52:48 (asr.sh:564:main) Skipped stages:  9 14 15 
2024-08-09T18:52:48 (asr.sh:877:main) Stage 5: Generate token_list from dump/raw/org/train/text using BPE
sentencepiece_trainer.cc(177) LOG(INFO) Running command: --input=data/token_list/bpe_unigram4234/train.txt --vocab_size=4234 --model_type=unigram --model_prefix=data/token_list/bpe_unigram4234/bpe --character_coverage=1.0 --input_sentence_size=100000000
sentencepiece_trainer.cc(77) LOG(INFO) Starts training with : 
trainer_spec {
  input: data/token_list/bpe_unigram4234/train.txt
  input_format: 
  model_prefix: data/token_list/bpe_unigram4234/bpe
  model_type: UNIGRAM
  vocab_size: 4234
  self_test_sample_size: 0
  character_coverage: 1
  input_sentence_size: 100000000
  shuffle_input_sentence: 1
  seed_sentencepiece_size: 1000000
  shrinking_factor: 0.75
  max_sentence_length: 4192
  num_threads: 16
  num_sub_iterations: 2
  max_sentencepiece_length: 16
  split_by_unicode_script: 1
  split_by_number: 1
  split_by_whitespace: 1
  split_digits: 0
  treat_whitespace_as_suffix: 0
  allow_whitespace_only_pieces: 0
  required_chars: 
  byte_fallback: 0
  vocabulary_output_piece_score: 1
  train_extremely_large_corpus: 0
  hard_vocab_limit: 1
  use_all_vocab: 0
  unk_id: 0
  bos_id: 1
  eos_id: 2
  pad_id: -1
  unk_piece: <unk>
  bos_piece: <s>
  eos_piece: </s>
  pad_piece: <pad>
  unk_surface:  ⁇ 
  enable_differential_privacy: 0
  differential_privacy_noise_level: 0
  differential_privacy_clipping_threshold: 0
}
normalizer_spec {
  name: nmt_nfkc
  add_dummy_prefix: 1
  remove_extra_whitespaces: 1
  escape_whitespaces: 1
  normalization_rule_tsv: 
}
denormalizer_spec {}
trainer_interface.cc(350) LOG(INFO) SentenceIterator is not specified. Using MultiFileSentenceIterator.
trainer_interface.cc(181) LOG(INFO) Loading corpus: data/token_list/bpe_unigram4234/train.txt
trainer_interface.cc(406) LOG(INFO) Loaded all 120098 sentences
trainer_interface.cc(422) LOG(INFO) Adding meta_piece: <unk>
trainer_interface.cc(422) LOG(INFO) Adding meta_piece: <s>
trainer_interface.cc(422) LOG(INFO) Adding meta_piece: </s>
trainer_interface.cc(427) LOG(INFO) Normalizing sentences...
trainer_interface.cc(536) LOG(INFO) all chars count=1850211
trainer_interface.cc(557) LOG(INFO) Alphabet size=4231
trainer_interface.cc(558) LOG(INFO) Final character coverage=1
trainer_interface.cc(590) LOG(INFO) Done! preprocessed 120098 sentences.
unigram_model_trainer.cc(146) LOG(INFO) Making suffix array...
unigram_model_trainer.cc(150) LOG(INFO) Extracting frequent sub strings...
unigram_model_trainer.cc(201) LOG(INFO) Initialized 384921 seed sentencepieces
trainer_interface.cc(596) LOG(INFO) Tokenizing input sentences with whitespace: 120098
trainer_interface.cc(607) LOG(INFO) Done! 113737
unigram_model_trainer.cc(491) LOG(INFO) Using 113737 sentences for EM training
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=169508 obj=60.2932 num_tokens=605625 num_tokens/piece=3.57284
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=154327 obj=57.151 num_tokens=608759 num_tokens/piece=3.9446
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=115321 obj=58.0022 num_tokens=643243 num_tokens/piece=5.57785
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=114890 obj=57.5467 num_tokens=644182 num_tokens/piece=5.60695
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=86118 obj=59.1864 num_tokens=685973 num_tokens/piece=7.9655
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=86071 obj=58.6914 num_tokens=686845 num_tokens/piece=7.97998
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=64542 obj=60.6321 num_tokens=730928 num_tokens/piece=11.3248
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=64533 obj=60.1345 num_tokens=731803 num_tokens/piece=11.34
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=48395 obj=62.3089 num_tokens=775418 num_tokens/piece=16.0227
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=48393 obj=61.8552 num_tokens=775965 num_tokens/piece=16.0347
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=36293 obj=64.1495 num_tokens=822372 num_tokens/piece=22.6592
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=36293 obj=63.7144 num_tokens=822793 num_tokens/piece=22.6708
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=27218 obj=66.1917 num_tokens=870513 num_tokens/piece=31.983
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=27218 obj=65.7753 num_tokens=871226 num_tokens/piece=32.0092
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=20412 obj=68.3477 num_tokens=920879 num_tokens/piece=45.1146
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=20412 obj=67.9286 num_tokens=921147 num_tokens/piece=45.1277
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=15309 obj=70.7446 num_tokens=974604 num_tokens/piece=63.6622
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=15309 obj=70.2996 num_tokens=975000 num_tokens/piece=63.688
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=11480 obj=73.3361 num_tokens=1034514 num_tokens/piece=90.1145
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=11479 obj=72.8184 num_tokens=1034923 num_tokens/piece=90.1579
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=8609 obj=76.2591 num_tokens=1105069 num_tokens/piece=128.362
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=8609 obj=75.6089 num_tokens=1105469 num_tokens/piece=128.409
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=6456 obj=80.114 num_tokens=1196553 num_tokens/piece=185.34
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=6456 obj=79.1711 num_tokens=1196816 num_tokens/piece=185.38
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=4842 obj=86.2754 num_tokens=1339006 num_tokens/piece=276.54
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=4842 obj=84.6424 num_tokens=1339362 num_tokens/piece=276.613
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=0 size=4657 obj=85.9046 num_tokens=1365198 num_tokens/piece=293.15
unigram_model_trainer.cc(507) LOG(INFO) EM sub_iter=1 size=4657 obj=85.6287 num_tokens=1365198 num_tokens/piece=293.15
trainer_interface.cc(685) LOG(INFO) Saving model: data/token_list/bpe_unigram4234/bpe.model
trainer_interface.cc(697) LOG(INFO) Saving vocabs: data/token_list/bpe_unigram4234/bpe.vocab
2024-08-09T18:52:58 (asr.sh:1809:main) Successfully finished. [elapsed=10s]
第 6-9 阶段：与语言建模相关的阶段
第 10 阶段：ASR 收集统计信息
# 第 10 阶段：ASR 收集统计信息：train_set=dump/raw/train，valid_set=dump/raw/dev
# 此阶段需要设置asr.sh中asr_config=asr_config=conf/train_asr_conformer.yaml
./asr.sh --stage 10 --stop_stage 10 --train_set train --valid_set dev --test_sets "dev test" --asr_config conf/train_asr_conformer.yaml

2024-08-09T18:56:13 (asr.sh:321:main) Info: The valid_set 'dev' is included in the test_sets. '--eval_valid_set true' is set and 'dev' is removed from the test_sets
2024-08-09T18:56:13 (asr.sh:564:main) Skipped stages:  9 14 15 
2024-08-09T18:56:13 (asr.sh:1189:main) Stage 10: ASR collect stats: train_set=dump/raw/train, valid_set=dump/raw/dev
2024-08-09T18:56:13 (asr.sh:1240:main) Generate 'exp/asr_stats_raw_bpe4234/run.sh'. You can resume the process from stage 10 using this script
2024-08-09T18:56:13 (asr.sh:1244:main) ASR collect-stats started... log: 'exp/asr_stats_raw_bpe4234/logdir/stats.*.log'
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/aggregate_stats_dirs.py --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.1 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.2 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.3 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.4 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.5 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.6 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.7 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.8 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.9 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.10 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.11 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.12 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.13 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.14 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.15 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.16 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.17 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.18 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.19 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.20 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.21 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.22 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.23 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.24 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.25 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.26 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.27 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.28 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.29 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.30 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.31 --input_dir exp/asr_stats_raw_bpe4234/logdir/stats.32 --output_dir exp/asr_stats_raw_bpe4234
2024-08-09T19:01:59 (asr.sh:1809:main) Successfully finished. [elapsed=346s]
第 11 阶段：ASR 训练
# 第 11 阶段：ASR 训练：train_set=dump/raw/train，valid_set=dump/raw/dev
# 更改train_asr_conformer.yaml中max_epoch: 50，其他默认设置
export CUDA_VISIBLE_DEVICES=1,3,4,5
./asr.sh --stage 11 --stop_stage 11 --train_set train --valid_set dev --test_sets "dev test" --asr_config conf/train_asr_conformer.yaml --ngpu 4

2024-08-10T09:27:53 (asr.sh:321:main) Info: The valid_set 'dev' is included in the test_sets. '--eval_valid_set true' is set and 'dev' is removed from the test_sets
2024-08-10T09:27:53 (asr.sh:564:main) Skipped stages:  9 14 15 
2024-08-10T09:27:53 (asr.sh:1308:main) Stage 11: ASR Training: train_set=dump/raw/train, valid_set=dump/raw/dev
2024-08-10T09:27:53 (asr.sh:1407:main) Generate 'exp/asr_train_asr_conformer_raw_bpe4234/run.sh'. You can resume the process from stage 11 using this script
2024-08-10T09:27:53 (asr.sh:1411:main) ASR training started... log: 'exp/asr_train_asr_conformer_raw_bpe4234/train.log'
2024-08-10 09:27:53,791 (launch:94) INFO: /home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/launch.py --cmd 'run.pl --name exp/asr_train_asr_conformer_raw_bpe4234/train.log' --log exp/asr_train_asr_conformer_raw_bpe4234/train.log --ngpu 4 --num_nodes 1 --init_file_prefix exp/asr_train_asr_conformer_raw_bpe4234/.dist_init_ --multiprocessing_distributed true -- python3 -m espnet2.bin.asr_train --use_preprocessor true --bpemodel data/token_list/bpe_unigram4234/bpe.model --token_type bpe --token_list data/token_list/bpe_unigram4234/tokens.txt --non_linguistic_symbols none --cleaner none --g2p none --valid_data_path_and_name_and_type dump/raw/dev/wav.scp,speech,sound --valid_shape_file exp/asr_stats_raw_bpe4234/valid/speech_shape --resume true --ignore_init_mismatch false --fold_length 80000 --output_dir exp/asr_train_asr_conformer_raw_bpe4234 --config conf/train_asr_conformer.yaml --frontend_conf fs=16k --normalize=global_mvn --normalize_conf stats_file=exp/asr_stats_raw_bpe4234/train/feats_stats.npz --train_data_path_and_name_and_type dump/raw/train/wav.scp,speech,sound --train_shape_file exp/asr_stats_raw_bpe4234/train/speech_shape --fold_length 150 --train_data_path_and_name_and_type dump/raw/train/text,text,text --train_shape_file exp/asr_stats_raw_bpe4234/train/text_shape.bpe --valid_data_path_and_name_and_type dump/raw/dev/text,text,text --valid_shape_file exp/asr_stats_raw_bpe4234/valid/text_shape.bpe
2024-08-10 09:27:53,806 (launch:237) INFO: single-node with 4gpu on distributed mode
2024-08-10 09:27:53,809 (launch:348) INFO: log file: exp/asr_train_asr_conformer_raw_bpe4234/train.log
2024-08-10T15:46:41 (asr.sh:1809:main) Successfully finished. [elapsed=22728s]

tensorboard可视化

由于我使用的是服务器，无法生成链接，同wenet一样，将训练生成的tensorboard文件，拿到本机上运行查看效果。

tensorboard --logdir tensorboard/train/ --port 12598 --bind_all

在这里插入图片描述

tensorboard --logdir tensorboard/valid/ --port 12598 --bind_all

在这里插入图片描述

第 12 阶段：解码
export CUDA_VISIBLE_DEVICES=1,3,4,5
./asr.sh --use_lm false --gpu_inference true --inference_nj 4 --stage 12 --stop_stage 12 --train_set train --valid_set dev --test_sets "dev test" --asr_exp exp/asr_train_asr_conformer_raw_bpe4234 --inference_config conf/decode_asr_transformer.yaml --ngpu 4

2024-08-11T09:02:43 (asr.sh:283:main) ./asr.sh --use_lm false --gpu_inference true --inference_nj 4 --stage 12 --stop_stage 12 --train_set train --valid_set dev --test_sets dev test --asr_exp exp/asr_train_asr_conformer_raw_bpe4234 --inference_config conf/decode_asr_transformer.yaml --ngpu 42024-08-11T09:02:43 (asr.sh:321:main) Info: The valid_set 'dev' is included in the test_sets. '--eval_valid_set true' is set and 'dev' is removed from the test_sets
2024-08-11T09:02:43 (asr.sh:564:main) Skipped stages:  6 7 8 9 14 15 
2024-08-11T09:02:43 (asr.sh:1480:main) Stage 12: Decoding: training_dir=exp/asr_train_asr_conformer_raw_bpe42342024-08-11T09:02:43 (asr.sh:1508:main) Generate 'exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/run.sh'. You can resume the process from stage 12 using this script
2024-08-11T09:02:43 (asr.sh:1573:main) Decoding started... log: 'exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/org/dev/logdir/asr_inference.*.log'
2024-08-11T11:58:37 (asr.sh:1589:main) Calculating RTF & latency... log: 'exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/org/dev/logdir/calculate_rtf.log'
2024-08-11T11:58:40 (asr.sh:1573:main) Decoding started... log: 'exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/test/logdir/asr_inference.*.log'
2024-08-11T13:35:20 (asr.sh:1589:main) Calculating RTF & latency... log: 'exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/test/logdir/calculate_rtf.log'
2024-08-11T13:35:21 (asr.sh:1809:main) Successfully finished. [elapsed=16358s]
第 13 阶段：得分
./asr.sh --stage 13 --stop_stage 13 --train_set train --valid_set dev --test_sets "dev test" --use_lm false --asr_exp exp/asr_train_asr_conformer_raw_bpe4234 --inference_config decode_asr_transformer.yaml

2024-08-11T14:41:53 (asr.sh:321:main) Info: The valid_set 'dev' is included in the test_sets. '--eval_valid_set true' is set and 'dev' is removed from the test_sets
2024-08-11T14:41:54 (asr.sh:564:main) Skipped stages:  6 7 8 9 14 15 
2024-08-11T14:41:54 (asr.sh:1621:main) Stage 13: Scoring
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --cleaner none --token_type char --non_linguistic_symbols none --remove_non_linguistic_symbols true
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --token_type char --non_linguistic_symbols none --remove_non_linguistic_symbols true --cleaner none
2024-08-11T14:42:36 (asr.sh:1711:main) Write cer result in exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/org/dev/score_cer/result.txt
|   SPKR      |   # Snt        # Wrd    |   Corr         Sub         Del          Ins         Err       S.Err    |
|   Sum/Avg   |  14326        205341    |   94.8         5.1         0.1          0.1         5.3        38.9    |
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --cleaner none --token_type word --non_linguistic_symbols none --remove_non_linguistic_symbols true
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --token_type word --non_linguistic_symbols none --remove_non_linguistic_symbols true --cleaner none
2024-08-11T14:43:04 (asr.sh:1711:main) Write wer result in exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/org/dev/score_wer/result.txt
|   SPKR      |    # Snt      # Wrd    |   Corr          Sub         Del         Ins          Err       S.Err    |
|   Sum/Avg   |   14326       14326    |   61.1         38.9         0.0         0.0         38.9        38.9    |
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --cleaner none --token_type bpe --bpemodel data/token_list/bpe_unigram4234/bpe.model
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --token_type bpe --bpemodel data/token_list/bpe_unigram4234/bpe.model --cleaner none
2024-08-11T14:43:48 (asr.sh:1711:main) Write ter result in exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/org/dev/score_ter/result.txt
|   SPKR      |   # Snt        # Wrd    |   Corr         Sub         Del          Ins         Err       S.Err    |
|   Sum/Avg   |  14326        219662    |   95.2         4.7         0.1          0.1         4.9        38.9    |
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --cleaner none --token_type char --non_linguistic_symbols none --remove_non_linguistic_symbols true
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --token_type char --non_linguistic_symbols none --remove_non_linguistic_symbols true --cleaner none
2024-08-11T14:44:09 (asr.sh:1711:main) Write cer result in exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/test/score_cer/result.txt
|   SPKR      |   # Snt       # Wrd    |   Corr         Sub        Del         Ins         Err       S.Err    |
|   Sum/Avg   |   7176       104765    |   94.5         5.4        0.2         0.1         5.7        40.2    |
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --cleaner none --token_type word --non_linguistic_symbols none --remove_non_linguistic_symbols true
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --token_type word --non_linguistic_symbols none --remove_non_linguistic_symbols true --cleaner none
2024-08-11T14:44:29 (asr.sh:1711:main) Write wer result in exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/test/score_wer/result.txt
|   SPKR      |   # Snt      # Wrd    |   Corr         Sub         Del         Ins         Err       S.Err    |
|   Sum/Avg   |   7176        7176    |   59.8        40.2         0.0         0.0        40.2        40.2    |
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --cleaner none --token_type bpe --bpemodel data/token_list/bpe_unigram4234/bpe.model
/home/lnj524/miniconda3/envs/espnet/bin/python3 /s6home/lnj524/module/espnet/espnet2/bin/tokenize_text.py -f 2- --input - --output - --token_type bpe --bpemodel data/token_list/bpe_unigram4234/bpe.model --cleaner none
2024-08-11T14:44:46 (asr.sh:1711:main) Write ter result in exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave/test/score_ter/result.txt
|   SPKR      |   # Snt       # Wrd    |   Corr         Sub        Del         Ins         Err       S.Err    |
|   Sum/Avg   |   7176       111941    |   94.8         5.0        0.1         0.1         5.3        40.2    |

## exp/asr_train_asr_conformer_raw_bpe4234
### WER

|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|---|---|---|---|---|---|---|---|---|
|decode_asr_transformer_asr_model_valid.acc.ave/test|7176|7176|59.8|40.2|0.0|0.0|40.2|40.2|

### CER

|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|---|---|---|---|---|---|---|---|---|
|decode_asr_transformer_asr_model_valid.acc.ave/test|7176|104765|94.5|5.4|0.2|0.1|5.7|40.2|

### TER

|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|---|---|---|---|---|---|---|---|---|
|decode_asr_transformer_asr_model_valid.acc.ave/test|7176|111941|94.8|5.0|0.1|0.1|5.3|40.2|

## exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave
### WER

|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|---|---|---|---|---|---|---|---|---|
|org/dev|14326|14326|61.1|38.9|0.0|0.0|38.9|38.9|

### CER

|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|---|---|---|---|---|---|---|---|---|
|org/dev|14326|205341|94.8|5.1|0.1|0.1|5.3|38.9|

### TER

|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|---|---|---|---|---|---|---|---|---|
|org/dev|14326|219662|95.2|4.7|0.1|0.1|4.9|38.9|

2024-08-11T14:44:51 (asr.sh:1809:main) Successfully finished. [elapsed=178s]

exp/asr_train_asr_conformer_raw_bpe4234

WER

datasetSntWrdCorrSubDelInsErrS.Err
decode_asr_transformer_asr_model_valid.acc.ave/test7176717659.840.20.00.040.240.2

CER

datasetSntWrdCorrSubDelInsErrS.Err
decode_asr_transformer_asr_model_valid.acc.ave/test717610476594.55.40.20.15.740.2

TER

datasetSntWrdCorrSubDelInsErrS.Err
decode_asr_transformer_asr_model_valid.acc.ave/test717611194194.85.00.10.15.340.2

exp/asr_train_asr_conformer_raw_bpe4234/decode_asr_transformer_asr_model_valid.acc.ave

WER

datasetSntWrdCorrSubDelInsErrS.Err
org/dev143261432661.138.90.00.038.938.9

CER

datasetSntWrdCorrSubDelInsErrS.Err
org/dev1432620534194.85.10.10.15.338.9

TER

datasetSntWrdCorrSubDelInsErrS.Err
org/dev1432621966295.24.70.10.14.938.9
@李思成
关注
  • 24
    点赞
  • 踩
  • 16
    收藏
    觉得还不错? 一键收藏
  • 打赏
    打赏
  • 0
    评论
专栏目录
kaldi环境配置与aishell实践
李思成的博客
08-11 378
如果某些库未安装，脚本会给出提示，你只需根据提示安装缺失的库即可。特别需要注意的是，Kaldi默认使用Intel MKL作为线性代数库。在编译Kaldi之前，需要检查和安装Kaldi依赖的系统开发库，包括g++、LLVM、Clang、zlib、python、gawk、perl、wget、git、libtool等。如果能够访问GitHub，ATLAS headers、OpenFst、SCTK、sph2pipe和CUB将会被自动下载、编译和安装。这是三个不同的语言模型工具，不同的示例使用不同的工具。
语音合成 GAN Vocoders 总览
小湉湉的博客
10-24 847
语音合成 GAN Vocoders 总览
Hidden Marlov Model（HMM）
qq_58317297的博客
08-14 271
1、Tandem：语音辨识HMM系统不变，唯一变化的是使用DNN（state classifer）给我们比较好的state（从声学特征到state），其他不变。c、如果想要更精确，可以使用训练出来的更精确的classifier作为寻找state与声学特征对应关系的Model，训练处DNN2。1、将声学特征设为X，经过语音识别得到的tokens设为Y，目标是找到通过X得到Y的最大概率，可以通过概率公式改变为。1、Transition Probability：从一个state转化为别的state的可能性。
智能语音识别技术在无人驾驶领域的应用案例
biaobei的博客
08-14 877
随着无人驾驶技术的进步与发展，越来越多的企业、创业者注意到无人驾驶领域潜藏的巨大市场经济价值，越来越多的企业和创业者进入无人驾驶领域，以近期业内关注的萝卜快跑为例，其在武汉地区的成功推广与落地预示着无人驾驶在网约车领域具有极大的可行性和经济价值。萝卜快跑无人驾驶的成功试投，也预示着智能语音交互在无人驾驶领域的成功应用。
语音播报加入预警系统
qq_57484399的博客
08-14 532
我们前一章, 已经把jq8900-16p模块, 单独进行了测试, 可以通过发送命令, 让模块播报设定好的声音。
动手研发实时口译系统
08-14 170
使用Whisper模型进行语音翻译（Speech-to-Speech Translation，STST）的过程。接下啦，我们使用Gradio库创建一个STST（Speech-to-Speech Translation）将英文文本转换成英文语音的过程，即文本到语音（Text-to-Speech，TTS）的转换。这里使用了专门针对英文TTS训练的SpeechT5模型。
win10配置pytorch环境+CUDA安装
SevenBerry的博客
08-12 601
打开每一项，将。
PyTorch FP16模型转换
weixin_67051070的博客
08-15 839
在使用 MNN 进行 INT8 量化推理时，你不需要在预处理阶段将图像转换为 INT8 格式。相反，你应该按照常规的图像预处理步骤，将图像转换为浮点数格式，然后将其输入到量化模型中。MNN 推理引擎会自动处理输入数据的量化。PyTorch：需要手动将输入数据转换为相应的推理格式（如 FP16 或 INT8）。MNN：不需要手动将输入数据转换为 INT8 格式，MNN 推理引擎会自动处理输入数据的量化。你只需按照常规的图像预处理步骤，将图像转换为浮点数格式即可。
PyTorch--残差网络（ResNet）在CIFAR-10数据集进行图像分类
最新发布
wumingzei的博客
08-15 604
定义一个3x3的卷积层。# ...定义残差网络中的残差块，包含两个卷积层和批量归一化层。# ...定义ResNet模型，使用残差块构建多个层。定义交叉熵损失函数和Adam优化器。
pytorch中的归一化函数
ym62033的博客
08-15 99
pytorch中的归一化函数_pytorch归一化-CSDN博客
模型训练坎坷路--逐步提升模型准确率从40%到90%+
m0_56243424的博客
08-15 605
剪枝实验的基准得是一个微调好的准确率保持较高的模型，按照我之前yolov5的经验来看，这还不简单?选模型+合适的数据集，20轮就OK!我计划使用 ResNet-18模型加上Cifar-10数据集，训练轮次100轮，准确度达到95%我就立马停下。可是过程却不大对劲，都20轮了我的准确率还是40%上下窜，一点进步的痕迹都没有，于是，果断按下停止键…from torchvision import models # 引入torchvision.modelstry:def train(
【pytorch学习】transforms与数据集的使用
qq_45059937的博客
08-15 297
新建一个名为test_dataset_transforms.py的文件由运行结果可以知道， 当前图片是PIL类型，现需要将其转为tensor类型，使用torch.ToTensor()展示图片。
pytorch学习（十四）：完整的模型训练套路
weixin_52307528的博客
08-12 319
以简单数据集CIFAR10为例，本次训练以GPU训练为例。讲解一下这三行代码的流程，首先optimizer.zero_grad() 将梯度清零，也就是每一层里的weight的grad里的参数清零，执行loss.backward()以后，grad里存储的就是反向传播的参数，optimizer.step()执行，将grad的值对参数进行更新。当你将代码块放在之下时，PyTorch会自动将该代码块中的所有操作设置为不计算梯度。
PyTorch模型转ONNX的3种方法
新缸中之脑
08-13 742
ONNX（开放神经网络交换）是一种用于表示机器学习模型的开放标准和格式。ONNX 定义了一组通用运算符和一种通用文件格式，用于表示各种框架中的深度学习模型，包括 PyTorch 和 TensorFlow。当将模型导出为 ONNX 格式时，这些运算符用于构建计算图（通常称为中间表示），该图表示数据通过神经网络的流动。重要提示： ONNX 不是运行时 ONNX 只是可以与 ONNX Runtime 等运行时一起使用的表示。你可以在此处找到受支持的加速器列表。
详细介绍Pytorch基于GPU训练的一般套路（device)
hello！
08-14 676
检查 GPU 可用性并选择设备。将模型移动到 GPU。在训练和评估循环中，将数据移动到 GPU。在保存和加载模型时，确保模型参数在 GPU 上。在 PyTorch 中，使用 GPU 进行模型训练可以显著提高训练速度，尤其是在处理大规模数据集和复杂模型时。以下是基于 GPU 训练的一般套路，包括数据准备、模型定义、损失函数和优化器的选择、训练循环、评估和测试，以及模型保存和加载。定义一个神经网络模型，通常继承自，并在__init__方法中定义网络层，在forward方法中定义前向传播过程。
什么是张量
2301_76924624的博客
08-14 870
学习使用pytorc库进行深度学习网络搭建时，张量这个词总是不定时会出现。其实，Pytorch中的所有操作都是在张量的基础上进行的，今天就来了解张量到底是什么由PyTorch官网官网介绍可知，一个Tensor是一个包含单一数据类型的多维矩阵。
动手学深度学习7.3 网络中的网络（NiN）-笔记&练习（PyTorch）
scdifsn的博客
08-12 814
李沐老师《动手学深度学习（PyTorch版）》7.3 网络中的网络（NiN）-笔记&练习详解
PyTorch--卷积神经网络（CNN）模型实现手写数字识别
wumingzei的博客
08-14 897
今天要介绍的这段代码是一个使用PyTorch框架实现的卷积神经网络（CNN）模型，用于对MNIST数据集进行分类的示例。MNIST数据集是手写数字识别领域的一个标准数据集，包含0到9的灰度图像。导入必要的库：导入PyTorch、PyTorch神经网络模块、torchvision（用于处理图像数据集）和transforms（用于图像预处理）。设备配置：设置模型运行的设备，优先使用GPU（如果可用），否则使用CPU。超参数设置：定义了训练迭代的轮数(num_epochs)、类别数()、批次大小(
Learning Multi-dimensional Human Preference for Text-to-Image Generation
kebijuelun的博客
08-13 829
2024 年快手提出的文生图评测方案，提出了多维度偏好评分（MPS），这是首个用于评估文本生成图像模型的多维度偏好评分模型。
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

@李思成

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付：¥1
获取中
扫码支付

您的余额不足，请更换扫码支付或充值

打赏作者

实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值