-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcommand.txt
89 lines (73 loc) · 3.15 KB
/
command.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
# 1. 安装依赖
pip install -r requirements.txt
pip install monotonic-align
pip install textparser-${lastest-version}.tar.gz
# 2. 数据准备
# 2.1 生成文本vector，下面操作会在.wav同目录下生成.vec256文件
cd /nfs/yuanwuwen/tts/data
cd guiji_230810
cat txt/all523.txt | text-parser -o wav
cd ../dataset
cat txt/all.txt | text-parser -o wav
cd ../s684
cat txt/s684.lab | text-vectorization -o wav
cd ../cmu_arctic
cat txt/all.txt | text-parser -o wav
cd ../cmu_arctic
cat all-108.txt | text-parser -o wav
# 2.2 提取情感向量，会在.wav同目录下生成.emo文件
cd ./VITS/emotional-vits3/
find /nfs/yuanwuwen/tts/data/dataset -name "*.wav" > files.scp
find /nfs/yuanwuwen/tts/data/guiji_230810 -name "*.wav" >> files.scp
find /nfs/yuanwuwen/tts/data/s684 -name "*.wav" >> files.scp
find /nfs/yuanwuwen/tts/data/cmu_arctic -name "*.wav" >> files.scp
find /nfs/yuanwuwen/tts/data/VCTK-Corpus -name "*.wav" >> files.scp
python toolkits/extract_emotion2.py --scp files.scp
# 上面会用到情感预训练模型: /nfs/yuanwuwen/tts/w2v2-L-robust-12.6bc4a7fd-1.1.0.zip，需要将该文件复制到~/.cache/audeer/cache/
rm -f files.scp
# 2.3 生成filelist
cd ./VITS/emotional-vits3/
# 参考filelists/s1325.scp中的格式
# 格式：vecfn|wavfn|emofn|spkid
# 2.4 配置configs/s1325.scp中的"training_files"和"validation_files"
# 3. training
# 3.1 training from scratch
# !!!注意：如果数据量比较大可能出现无法收敛的情况。解决方法是可以先使用100+hr的数据预训练10k steps，大概dur_loss下降到小于1.0
CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py -c configs/s1322.json -m s1322
# 也可切换鉴别式和stft loss继续训练，下面这步也可以不做。
#CUDA_VISIBLE_DEVICES=0,1,2,3 python train_stft.py -c configs/s1325.json -m s1325-stft-from-s1322 --ckptG logs/s1322/G_975000.pth
# 区别："train.py"使用mel_loss+adv_loss+fm_loss；"train_stft.py"使用stft_loss+adv_loss。并且鉴别式不一样。
# 3.2 finetune from pretrained
# 比如微调刘润liurun
python train.py -a -d -c configs/liurun.json -m liurun --ckptG logs/s1322/G_975000.pth --ckptD logs/s1322/D_975000.pth
!!!!!!!!!!!!!!!!!!!!
导出emo，# 以刘润liurun模型为例
# 1. 导出模型
python export.py --greedy 3 --checkpoint logs/liurun/ --outdir ./checkpoint/liurun
# 2. 聚类情感向量
# 每个发音人都有自己的emo聚类，根据每个发音人风格和录音数目，选择K的大小
K=7
spkid=166
cut -d "|" -f 3 filelists/liurun.scp > emo.scp
python toolkits/cluster_emotion.py $K emo.scp ./checkpoint/${spkid}.emo
# 3. 推理
#python vits_wrap.py --help
python vits_wrap.py -c ./checkpoint/liurun/checkpoint.pth -i ${spkid} -e "${spkid}:0" -t your/input.txt -o your/output/dir -n filename.wav -d "cuda:0"
!!!!!!!!!!!!!!!!!!!!!
# 1. 安装依赖
pip install -r requirments.txt
pip install textparser-${lastest-version}.tar.gz
pip install fbandext-${lastest-version}.tar.gz
# 2. 推理时需要的py文件
attentions.py
commons.py
export.py
infer.py
__init__.py
models.py
modules.py
utils.py
version.py
vits_wrap.py
# 3. 备份的模型位置如下，另外@吴海娥(黄蓉) 也有备份
ls -hl /nfs/yuanwuwen/tts/VITS/intergration_package