-
Notifications
You must be signed in to change notification settings - Fork 46
/
webui_train.py
136 lines (108 loc) · 6.02 KB
/
webui_train.py
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
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
import json
import os
import gradio as gr
import subprocess
from pathlib import Path
def preprocess(train_input_path, val_input_path, output_path, pre_model_path):
for state, input_path in zip(['train', 'val'], [train_input_path, val_input_path]):
temp1 = Path(output_path)/state/'temp1'
temp2 = Path(output_path)/state/'temp2'
try:
temp1.mkdir(parents=True)
temp2.mkdir(parents=True)
except Exception as e:
pass
print(input_path)
print(temp1)
subprocess.run([r'.\py311\python.exe', 'local/prepare_data.py',
'--src_dir', input_path,
'--des_dir', str(temp1)])
print("第一步结束")
subprocess.run([r'.\py311\python.exe', 'tools/extract_embedding.py',
'--dir', str(temp1),
'--onnx_path', "pretrained_models/CosyVoice-300M/campplus.onnx"
])
print("第二步结束")
subprocess.run([r'.\py311\python.exe', 'tools/extract_speech_token.py',
'--dir', str(temp1),
'--onnx_path', "pretrained_models/CosyVoice-300M/speech_tokenizer_v1.onnx"
])
print("第三步结束")
subprocess.run([r'.\py311\python.exe', 'tools/make_parquet_list.py',
'--num_utts_per_parquet', '100',
'--num_processes', '1',
'--src_dir', str(temp1),
'--des_dir', str(temp2),
])
return f'{state} make parquet list done!'
def refresh_voice(output_path):
content = (Path(output_path)/'train'/'temp1'/'utt2spk').read_text()
voices = []
for item in content.split('\n'):
voices.append(item.split(' ')[0])
return gr.Dropdown(choices=voices)
def train(output_path, pre_model_path):
train_list = os.path.join(output_path, 'train', 'temp2', 'data.list')
val_list = os.path.join(output_path, 'val', 'temp2', 'data.list')
model_dir = Path(output_path)/'models'
model_dir.mkdir(exist_ok=True, parents=True)
cmd = rf".\py311\Scripts\torchrun.exe --nnodes 1 --nproc_per_node 1 --rdzv_id 1986 --rdzv_backend c10d --rdzv_endpoint localhost:0 cosyvoice/bin/train.py --train_engine torch_ddp --config conf/cosyvoice.yaml --model llm --checkpoint pretrained_models/CosyVoice-300M/llm.pt --ddp.dist_backend gloo --num_workers 1 --prefetch 100 --pin_memory --train_data {train_list} --cv_data {val_list} --model_dir {model_dir} --tensorboard_dir {model_dir}"
print(cmd)
res = subprocess.Popen(cmd)
res.wait()
# subprocess.run([r'.\py311\torchrun', '--nnodes', '1', '--nproc_per_node', '1', '--rdzv_id', '1986',
# '--rdzv_backend', "c10d", '--rdzv_endpoint', "localhost:0",
# 'cosyvoice/bin/train.py','--train_engine','torch_ddp','--config','conf/cosyvoice.yaml',
# '--model','llm', '--checkpoint', os.path.join(pre_model_path, 'llm.pt'),
# '--ddp.dist_backend', 'nccl', '--num_workers', '1', '--prefetch', '100','--pin_memory',
# '--deepspeed_config', './conf/ds_stage2.json', '--deepspeed.save_states', 'model+optimizer',
# '--train_data', train_list, '--cv_data', val_list,
# '--model_dir', str(model_dir), '--tensorboard_dir', str(model_dir),
# ])
return 'Train done!'
def inference(mode, output_path, epoch, pre_model_path, text, voice):
train_list = os.path.join(output_path, 'train', 'temp2', 'data.list')
utt2data_list = Path(train_list).with_name('utt2data.list')
llm_model = os.path.join(output_path, 'models', f'epoch_{epoch}_whole.pt')
flow_model = os.path.join(pre_model_path, 'flow.pt')
hifigan_model = os.path.join(pre_model_path, 'hift.pt')
res_dir = Path(output_path)/'outputs'
res_dir.mkdir(exist_ok=True, parents=True)
json_path = str(Path(res_dir)/'tts_text.json')
with open(json_path, 'wt', encoding='utf-8') as f:
json.dump({voice:[text]}, f)
subprocess.run([r'.\pyr11\python.exe', 'cosyvoice/bin/inference.py',
'--mode', mode,
'--gpu', '0', '--config', 'conf/cosyvoice.yaml',
'--prompt_data', train_list,
'--prompt_utt2data', str(utt2data_list),
'--tts_text', json_path,
'--llm_model', llm_model,
'--flow_model', flow_model,
'--hifigan_model', hifigan_model,
'--result_dir', str(res_dir)])
output_path = str(Path(res_dir)/f'{voice}_0.wav')
return output_path
with gr.Blocks() as demo:
pretrained_model_path = gr.Text('./pretrained_models/CosyVoice-300M', label='Pretrained model dir')
output_dir = gr.Text(label='Output dir',value="./output")
with gr.Tab('Train'):
train_input_path = gr.Text(label='Train input path',value="./train")
val_input_path = gr.Text(label='Val input path',value="./val")
preprocess_btn = gr.Button('Preprocess', variant='primary')
train_btn = gr.Button('Train', variant='primary')
status = gr.Text(label='Status')
with gr.Tab('Inference'):
with gr.Row():
voices = gr.Dropdown(label='Voices')
refresh = gr.Button('Refresh voices', variant='primary')
mode = gr.Dropdown(choices=['sft', 'zero_shot'], label='Mode')
epoch = gr.Number(value=8, interactive=True, precision=0, label='Epochs')
text = gr.Text()
inference_btn = gr.Button('Inference', variant='primary')
out_audio = gr.Audio()
preprocess_btn.click(preprocess, inputs=[train_input_path, val_input_path, output_dir, pretrained_model_path], outputs=status)
train_btn.click(train, inputs=[output_dir, pretrained_model_path], outputs=status)
inference_btn.click(inference, inputs=[mode, output_dir, epoch, pretrained_model_path, text, voices], outputs=out_audio)
refresh.click(refresh_voice, inputs=output_dir, outputs=voices)
demo.launch(server_name='0.0.0.0',server_port=9883,inbrowser=True)