Upload 2 files
Browse files- dataset.py +66 -0
- tain.py +220 -0
dataset.py
ADDED
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
import sys
|
4 |
+
import tempfile
|
5 |
+
|
6 |
+
import librosa.display
|
7 |
+
import numpy as np
|
8 |
+
|
9 |
+
import os
|
10 |
+
import torch
|
11 |
+
import torchaudio
|
12 |
+
import traceback
|
13 |
+
from TTS.demos.xtts_ft_demo.utils.formatter import format_audio_list
|
14 |
+
from TTS.demos.xtts_ft_demo.utils.gpt_train import train_gpt
|
15 |
+
|
16 |
+
from TTS.tts.configs.xtts_config import XttsConfig
|
17 |
+
from TTS.tts.models.xtts import Xtts
|
18 |
+
|
19 |
+
|
20 |
+
def clear_gpu_cache():
|
21 |
+
# clear the GPU cache
|
22 |
+
if torch.cuda.is_available():
|
23 |
+
torch.cuda.empty_cache()
|
24 |
+
|
25 |
+
|
26 |
+
def preprocess_dataset(audio_path, language, out_path):
|
27 |
+
"""
|
28 |
+
Prepara los datos de audio para el entrenamiento del modelo.
|
29 |
+
|
30 |
+
Args:
|
31 |
+
audio_path (list): Lista de rutas de los archivos de audio.
|
32 |
+
language (str): Código del idioma del dataset.
|
33 |
+
out_path (str): Ruta de salida para el dataset procesado.
|
34 |
+
|
35 |
+
Returns:
|
36 |
+
tuple: Tupla con las rutas de los archivos CSV de entrenamiento y evaluación.
|
37 |
+
"""
|
38 |
+
out_path = os.path.join(out_path, "dataset")
|
39 |
+
os.makedirs(out_path, exist_ok=True)
|
40 |
+
train_meta, eval_meta, _ = format_audio_list(audio_path, target_language=language, out_path=out_path)
|
41 |
+
train_csv = os.path.join(out_path, "train.csv")
|
42 |
+
eval_csv = os.path.join(out_path, "eval.csv")
|
43 |
+
return train_csv, eval_csv
|
44 |
+
|
45 |
+
def main(dataset_path, output_path, language):
|
46 |
+
# Obtener información del usuario
|
47 |
+
audio_path = dataset_path #input("Ingresa la ruta de los archivos de audio (separados por espacio): ")
|
48 |
+
language = language #input("Ingresa el idioma del dataset: ")
|
49 |
+
out_path = output_path #input("Ingresa la ruta de salida para el dataset procesado: ")
|
50 |
+
|
51 |
+
# Prepara los datos
|
52 |
+
train_csv, eval_csv = preprocess_dataset(audio_path.split(), language, out_path)
|
53 |
+
|
54 |
+
print(f"Los archivos CSV se han creado en: {out_path}")
|
55 |
+
print(f"train.csv: {train_csv}")
|
56 |
+
print(f"eval.csv: {eval_csv}")
|
57 |
+
|
58 |
+
|
59 |
+
if __name__ == "__main__":
|
60 |
+
parser = argparse.ArgumentParser()
|
61 |
+
parser.add_argument("--dataset_path", type=str, required=True, help="Ruta del dataset de audio")
|
62 |
+
parser.add_argument("--output_path", type=str, required=True, help="Ruta de salida para el dataset procesado")
|
63 |
+
parser.add_argument("--language", type=str, required=True, help="Idioma del dataset")
|
64 |
+
args = parser.parse_args()
|
65 |
+
|
66 |
+
main(args.dataset_path, args.output_path, args.language)
|
tain.py
ADDED
@@ -0,0 +1,220 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
import sys
|
4 |
+
import tempfile
|
5 |
+
|
6 |
+
import librosa.display
|
7 |
+
import numpy as np
|
8 |
+
|
9 |
+
import os
|
10 |
+
import torch
|
11 |
+
import torchaudio
|
12 |
+
import traceback
|
13 |
+
from TTS.demos.xtts_ft_demo.utils.formatter import format_audio_list
|
14 |
+
from TTS.demos.xtts_ft_demo.utils.gpt_train import train_gpt
|
15 |
+
|
16 |
+
from TTS.tts.configs.xtts_config import XttsConfig
|
17 |
+
from TTS.tts.models.xtts import Xtts
|
18 |
+
|
19 |
+
|
20 |
+
def clear_gpu_cache():
|
21 |
+
# clear the GPU cache
|
22 |
+
if torch.cuda.is_available():
|
23 |
+
torch.cuda.empty_cache()
|
24 |
+
|
25 |
+
XTTS_MODEL = None
|
26 |
+
def load_model(xtts_checkpoint, xtts_config, xtts_vocab):
|
27 |
+
global XTTS_MODEL
|
28 |
+
clear_gpu_cache()
|
29 |
+
if not xtts_checkpoint or not xtts_config or not xtts_vocab:
|
30 |
+
return "You need to run the previous steps or manually set the `XTTS checkpoint path`, `XTTS config path`, and `XTTS vocab path` fields !!"
|
31 |
+
config = XttsConfig()
|
32 |
+
config.load_json(xtts_config)
|
33 |
+
XTTS_MODEL = Xtts.init_from_config(config)
|
34 |
+
print("Loading XTTS model! ")
|
35 |
+
XTTS_MODEL.load_checkpoint(config, checkpoint_path=xtts_checkpoint, vocab_path=xtts_vocab, use_deepspeed=False)
|
36 |
+
if torch.cuda.is_available():
|
37 |
+
XTTS_MODEL.cuda()
|
38 |
+
|
39 |
+
print("Model Loaded!")
|
40 |
+
return "Model Loaded!"
|
41 |
+
|
42 |
+
def run_tts(lang, tts_text, speaker_audio_file):
|
43 |
+
if XTTS_MODEL is None or not speaker_audio_file:
|
44 |
+
return "You need to run the previous step to load the model !!", None, None
|
45 |
+
|
46 |
+
gpt_cond_latent, speaker_embedding = XTTS_MODEL.get_conditioning_latents(audio_path=speaker_audio_file, gpt_cond_len=XTTS_MODEL.config.gpt_cond_len, max_ref_length=XTTS_MODEL.config.max_ref_len, sound_norm_refs=XTTS_MODEL.config.sound_norm_refs)
|
47 |
+
out = XTTS_MODEL.inference(
|
48 |
+
text=tts_text,
|
49 |
+
language=lang,
|
50 |
+
gpt_cond_latent=gpt_cond_latent,
|
51 |
+
speaker_embedding=speaker_embedding,
|
52 |
+
temperature=XTTS_MODEL.config.temperature, # Add custom parameters here
|
53 |
+
length_penalty=XTTS_MODEL.config.length_penalty,
|
54 |
+
repetition_penalty=XTTS_MODEL.config.repetition_penalty,
|
55 |
+
top_k=XTTS_MODEL.config.top_k,
|
56 |
+
top_p=XTTS_MODEL.config.top_p,
|
57 |
+
)
|
58 |
+
|
59 |
+
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp:
|
60 |
+
out["wav"] = torch.tensor(out["wav"]).unsqueeze(0)
|
61 |
+
out_path = fp.name
|
62 |
+
torchaudio.save(out_path, out["wav"], 24000)
|
63 |
+
|
64 |
+
return "Speech generated !", out_path, speaker_audio_file
|
65 |
+
|
66 |
+
|
67 |
+
|
68 |
+
|
69 |
+
# define a logger to redirect
|
70 |
+
class Logger:
|
71 |
+
def __init__(self, filename="log.out"):
|
72 |
+
self.log_file = filename
|
73 |
+
self.terminal = sys.stdout
|
74 |
+
self.log = open(self.log_file, "w")
|
75 |
+
|
76 |
+
def write(self, message):
|
77 |
+
self.terminal.write(message)
|
78 |
+
self.log.write(message)
|
79 |
+
|
80 |
+
def flush(self):
|
81 |
+
self.terminal.flush()
|
82 |
+
self.log.flush()
|
83 |
+
|
84 |
+
def isatty(self):
|
85 |
+
return False
|
86 |
+
|
87 |
+
# redirect stdout and stderr to a file
|
88 |
+
sys.stdout = Logger()
|
89 |
+
sys.stderr = sys.stdout
|
90 |
+
|
91 |
+
|
92 |
+
# logging.basicConfig(stream=sys.stdout, level=logging.INFO)
|
93 |
+
import logging
|
94 |
+
logging.basicConfig(
|
95 |
+
level=logging.INFO,
|
96 |
+
format="%(asctime)s [%(levelname)s] %(message)s",
|
97 |
+
handlers=[
|
98 |
+
logging.StreamHandler(sys.stdout)
|
99 |
+
]
|
100 |
+
)
|
101 |
+
|
102 |
+
def read_logs():
|
103 |
+
sys.stdout.flush()
|
104 |
+
with open(sys.stdout.log_file, "r") as f:
|
105 |
+
return f.read()
|
106 |
+
|
107 |
+
|
108 |
+
if __name__ == "__main__":
|
109 |
+
|
110 |
+
parser = argparse.ArgumentParser(
|
111 |
+
description="""XTTS fine-tuning demo\n\n"""
|
112 |
+
"""
|
113 |
+
Example runs:
|
114 |
+
python3 TTS/demos/xtts_ft_demo/xtts_demo.py --port
|
115 |
+
""",
|
116 |
+
formatter_class=argparse.RawTextHelpFormatter,
|
117 |
+
)
|
118 |
+
parser.add_argument(
|
119 |
+
"--port",
|
120 |
+
type=int,
|
121 |
+
help="Port to run the gradio demo. Default: 5003",
|
122 |
+
default=5003,
|
123 |
+
)
|
124 |
+
parser.add_argument(
|
125 |
+
"--out_path",
|
126 |
+
type=str,
|
127 |
+
help="Output path (where data and checkpoints will be saved) Default: /tmp/xtts_ft/",
|
128 |
+
default="/tmp/xtts_ft/",
|
129 |
+
)
|
130 |
+
|
131 |
+
parser.add_argument(
|
132 |
+
"--num_epochs",
|
133 |
+
type=int,
|
134 |
+
help="Number of epochs to train. Default: 10",
|
135 |
+
default=10,
|
136 |
+
)
|
137 |
+
parser.add_argument(
|
138 |
+
"--batch_size",
|
139 |
+
type=int,
|
140 |
+
help="Batch size. Default: 4",
|
141 |
+
default=4,
|
142 |
+
)
|
143 |
+
parser.add_argument(
|
144 |
+
"--grad_acumm",
|
145 |
+
type=int,
|
146 |
+
help="Grad accumulation steps. Default: 1",
|
147 |
+
default=1,
|
148 |
+
)
|
149 |
+
parser.add_argument(
|
150 |
+
"--max_audio_length",
|
151 |
+
type=int,
|
152 |
+
help="Max permitted audio size in seconds. Default: 11",
|
153 |
+
default=11,
|
154 |
+
)
|
155 |
+
|
156 |
+
# Add the new arguments
|
157 |
+
parser.add_argument(
|
158 |
+
"--lang",
|
159 |
+
type=str,
|
160 |
+
help="Dataset Language",
|
161 |
+
default="en",
|
162 |
+
)
|
163 |
+
parser.add_argument(
|
164 |
+
"--train_csv",
|
165 |
+
type=str,
|
166 |
+
help="Path to the train CSV file",
|
167 |
+
required=True,
|
168 |
+
)
|
169 |
+
parser.add_argument(
|
170 |
+
"--eval_csv",
|
171 |
+
type=str,
|
172 |
+
help="Path to the eval CSV file",
|
173 |
+
required=True,
|
174 |
+
)
|
175 |
+
|
176 |
+
args = parser.parse_args()
|
177 |
+
|
178 |
+
# ... (rest of your code)
|
179 |
+
|
180 |
+
def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acumm, output_path, max_audio_length):
|
181 |
+
clear_gpu_cache()
|
182 |
+
if not train_csv or not eval_csv:
|
183 |
+
return "You need to run the data processing step or manually set `Train CSV` and `Eval CSV` fields !", "", "", "", ""
|
184 |
+
try:
|
185 |
+
# convert seconds to waveform frames
|
186 |
+
max_audio_length = int(max_audio_length * 22050)
|
187 |
+
config_path, original_xtts_checkpoint, vocab_file, exp_path, speaker_wav = train_gpt(language, num_epochs, batch_size, grad_acumm, train_csv, eval_csv, output_path=output_path, max_audio_length=max_audio_length)
|
188 |
+
except:
|
189 |
+
traceback.print_exc()
|
190 |
+
error = traceback.format_exc()
|
191 |
+
return f"The training was interrupted due an error !! Please check the console to check the full error message! \n Error summary: {error}", "", "", "", ""
|
192 |
+
|
193 |
+
# copy original files to avoid parameters changes issues
|
194 |
+
os.system(f"cp {config_path} {exp_path}")
|
195 |
+
os.system(f"cp {vocab_file} {exp_path}")
|
196 |
+
|
197 |
+
ft_xtts_checkpoint = os.path.join(exp_path, "best_model.pth")
|
198 |
+
print("Model training done!")
|
199 |
+
clear_gpu_cache()
|
200 |
+
return "Model training done!", config_path, vocab_file, ft_xtts_checkpoint, speaker_wav
|
201 |
+
|
202 |
+
# ... (rest of your code)
|
203 |
+
|
204 |
+
# The following section is the only part to be changed:
|
205 |
+
# It now directly calls the train_model function instead of using Gradio
|
206 |
+
|
207 |
+
if __name__ == "__main__":
|
208 |
+
# ... (argparse setup)
|
209 |
+
|
210 |
+
# Call the function directly
|
211 |
+
train_model(
|
212 |
+
args.lang,
|
213 |
+
args.train_csv,
|
214 |
+
args.eval_csv,
|
215 |
+
args.num_epochs,
|
216 |
+
args.batch_size,
|
217 |
+
args.grad_acumm,
|
218 |
+
args.out_path,
|
219 |
+
args.max_audio_length,
|
220 |
+
)
|