Spaces:
Running
Running
mrfakename
commited on
Commit
•
b6584c2
1
Parent(s):
0cc615c
Sync from GitHub repo
Browse filesThis Space is synced from the GitHub repo: https://github.com/SWivid/F5-TTS. Please submit contributions to the Space there
- app.py +2 -1
- src/f5_tts/api.py +2 -1
- src/f5_tts/infer/infer_cli.py +4 -2
- src/f5_tts/infer/utils_infer.py +4 -4
- src/f5_tts/model/trainer.py +42 -4
- src/f5_tts/train/finetune_cli.py +11 -0
- src/f5_tts/train/finetune_gradio.py +79 -0
- src/f5_tts/train/train.py +1 -0
app.py
CHANGED
@@ -37,7 +37,7 @@ from f5_tts.infer.utils_infer import (
|
|
37 |
save_spectrogram,
|
38 |
)
|
39 |
|
40 |
-
|
41 |
|
42 |
|
43 |
# load models
|
@@ -94,6 +94,7 @@ def infer(
|
|
94 |
ref_text,
|
95 |
gen_text,
|
96 |
ema_model,
|
|
|
97 |
cross_fade_duration=cross_fade_duration,
|
98 |
speed=speed,
|
99 |
show_info=show_info,
|
|
|
37 |
save_spectrogram,
|
38 |
)
|
39 |
|
40 |
+
vocoder = load_vocoder()
|
41 |
|
42 |
|
43 |
# load models
|
|
|
94 |
ref_text,
|
95 |
gen_text,
|
96 |
ema_model,
|
97 |
+
vocoder,
|
98 |
cross_fade_duration=cross_fade_duration,
|
99 |
speed=speed,
|
100 |
show_info=show_info,
|
src/f5_tts/api.py
CHANGED
@@ -47,7 +47,7 @@ class F5TTS:
|
|
47 |
self.load_ema_model(model_type, ckpt_file, vocab_file, ode_method, use_ema)
|
48 |
|
49 |
def load_vocoder_model(self, local_path):
|
50 |
-
self.
|
51 |
|
52 |
def load_ema_model(self, model_type, ckpt_file, vocab_file, ode_method, use_ema):
|
53 |
if model_type == "F5-TTS":
|
@@ -102,6 +102,7 @@ class F5TTS:
|
|
102 |
ref_text,
|
103 |
gen_text,
|
104 |
self.ema_model,
|
|
|
105 |
show_info=show_info,
|
106 |
progress=progress,
|
107 |
target_rms=target_rms,
|
|
|
47 |
self.load_ema_model(model_type, ckpt_file, vocab_file, ode_method, use_ema)
|
48 |
|
49 |
def load_vocoder_model(self, local_path):
|
50 |
+
self.vocoder = load_vocoder(local_path is not None, local_path, self.device)
|
51 |
|
52 |
def load_ema_model(self, model_type, ckpt_file, vocab_file, ode_method, use_ema):
|
53 |
if model_type == "F5-TTS":
|
|
|
102 |
ref_text,
|
103 |
gen_text,
|
104 |
self.ema_model,
|
105 |
+
self.vocoder,
|
106 |
show_info=show_info,
|
107 |
progress=progress,
|
108 |
target_rms=target_rms,
|
src/f5_tts/infer/infer_cli.py
CHANGED
@@ -113,7 +113,7 @@ wave_path = Path(output_dir) / "infer_cli_out.wav"
|
|
113 |
# spectrogram_path = Path(output_dir) / "infer_cli_out.png"
|
114 |
vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
|
115 |
|
116 |
-
|
117 |
|
118 |
|
119 |
# load models
|
@@ -175,7 +175,9 @@ def main_process(ref_audio, ref_text, text_gen, model_obj, remove_silence, speed
|
|
175 |
ref_audio = voices[voice]["ref_audio"]
|
176 |
ref_text = voices[voice]["ref_text"]
|
177 |
print(f"Voice: {voice}")
|
178 |
-
audio, final_sample_rate, spectragram = infer_process(
|
|
|
|
|
179 |
generated_audio_segments.append(audio)
|
180 |
|
181 |
if generated_audio_segments:
|
|
|
113 |
# spectrogram_path = Path(output_dir) / "infer_cli_out.png"
|
114 |
vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
|
115 |
|
116 |
+
vocoder = load_vocoder(is_local=args.load_vocoder_from_local, local_path=vocos_local_path)
|
117 |
|
118 |
|
119 |
# load models
|
|
|
175 |
ref_audio = voices[voice]["ref_audio"]
|
176 |
ref_text = voices[voice]["ref_text"]
|
177 |
print(f"Voice: {voice}")
|
178 |
+
audio, final_sample_rate, spectragram = infer_process(
|
179 |
+
ref_audio, ref_text, gen_text, model_obj, vocoder, speed=speed
|
180 |
+
)
|
181 |
generated_audio_segments.append(audio)
|
182 |
|
183 |
if generated_audio_segments:
|
src/f5_tts/infer/utils_infer.py
CHANGED
@@ -29,9 +29,6 @@ _ref_audio_cache = {}
|
|
29 |
|
30 |
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
|
31 |
|
32 |
-
vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
|
33 |
-
|
34 |
-
|
35 |
# -----------------------------------------
|
36 |
|
37 |
target_sample_rate = 24000
|
@@ -263,6 +260,7 @@ def infer_process(
|
|
263 |
ref_text,
|
264 |
gen_text,
|
265 |
model_obj,
|
|
|
266 |
show_info=print,
|
267 |
progress=tqdm,
|
268 |
target_rms=target_rms,
|
@@ -287,6 +285,7 @@ def infer_process(
|
|
287 |
ref_text,
|
288 |
gen_text_batches,
|
289 |
model_obj,
|
|
|
290 |
progress=progress,
|
291 |
target_rms=target_rms,
|
292 |
cross_fade_duration=cross_fade_duration,
|
@@ -307,6 +306,7 @@ def infer_batch_process(
|
|
307 |
ref_text,
|
308 |
gen_text_batches,
|
309 |
model_obj,
|
|
|
310 |
progress=tqdm,
|
311 |
target_rms=0.1,
|
312 |
cross_fade_duration=0.15,
|
@@ -362,7 +362,7 @@ def infer_batch_process(
|
|
362 |
generated = generated.to(torch.float32)
|
363 |
generated = generated[:, ref_audio_len:, :]
|
364 |
generated_mel_spec = generated.permute(0, 2, 1)
|
365 |
-
generated_wave =
|
366 |
if rms < target_rms:
|
367 |
generated_wave = generated_wave * rms / target_rms
|
368 |
|
|
|
29 |
|
30 |
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
|
31 |
|
|
|
|
|
|
|
32 |
# -----------------------------------------
|
33 |
|
34 |
target_sample_rate = 24000
|
|
|
260 |
ref_text,
|
261 |
gen_text,
|
262 |
model_obj,
|
263 |
+
vocoder,
|
264 |
show_info=print,
|
265 |
progress=tqdm,
|
266 |
target_rms=target_rms,
|
|
|
285 |
ref_text,
|
286 |
gen_text_batches,
|
287 |
model_obj,
|
288 |
+
vocoder,
|
289 |
progress=progress,
|
290 |
target_rms=target_rms,
|
291 |
cross_fade_duration=cross_fade_duration,
|
|
|
306 |
ref_text,
|
307 |
gen_text_batches,
|
308 |
model_obj,
|
309 |
+
vocoder,
|
310 |
progress=tqdm,
|
311 |
target_rms=0.1,
|
312 |
cross_fade_duration=0.15,
|
|
|
362 |
generated = generated.to(torch.float32)
|
363 |
generated = generated[:, ref_audio_len:, :]
|
364 |
generated_mel_spec = generated.permute(0, 2, 1)
|
365 |
+
generated_wave = vocoder.decode(generated_mel_spec.cpu())
|
366 |
if rms < target_rms:
|
367 |
generated_wave = generated_wave * rms / target_rms
|
368 |
|
src/f5_tts/model/trainer.py
CHANGED
@@ -6,6 +6,7 @@ from tqdm import tqdm
|
|
6 |
import wandb
|
7 |
|
8 |
import torch
|
|
|
9 |
from torch.optim import AdamW
|
10 |
from torch.utils.data import DataLoader, Dataset, SequentialSampler
|
11 |
from torch.optim.lr_scheduler import LinearLR, SequentialLR
|
@@ -39,9 +40,11 @@ class Trainer:
|
|
39 |
max_grad_norm=1.0,
|
40 |
noise_scheduler: str | None = None,
|
41 |
duration_predictor: torch.nn.Module | None = None,
|
|
|
42 |
wandb_project="test_e2-tts",
|
43 |
wandb_run_name="test_run",
|
44 |
wandb_resume_id: str = None,
|
|
|
45 |
last_per_steps=None,
|
46 |
accelerate_kwargs: dict = dict(),
|
47 |
ema_kwargs: dict = dict(),
|
@@ -49,21 +52,25 @@ class Trainer:
|
|
49 |
):
|
50 |
ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
|
51 |
|
52 |
-
logger
|
|
|
53 |
print(f"Using logger: {logger}")
|
|
|
54 |
|
55 |
self.accelerator = Accelerator(
|
56 |
-
log_with=logger,
|
57 |
kwargs_handlers=[ddp_kwargs],
|
58 |
gradient_accumulation_steps=grad_accumulation_steps,
|
59 |
**accelerate_kwargs,
|
60 |
)
|
61 |
|
62 |
-
|
|
|
63 |
if exists(wandb_resume_id):
|
64 |
init_kwargs = {"wandb": {"resume": "allow", "name": wandb_run_name, "id": wandb_resume_id}}
|
65 |
else:
|
66 |
init_kwargs = {"wandb": {"resume": "allow", "name": wandb_run_name}}
|
|
|
67 |
self.accelerator.init_trackers(
|
68 |
project_name=wandb_project,
|
69 |
init_kwargs=init_kwargs,
|
@@ -81,11 +88,15 @@ class Trainer:
|
|
81 |
},
|
82 |
)
|
83 |
|
|
|
|
|
|
|
|
|
|
|
84 |
self.model = model
|
85 |
|
86 |
if self.is_main:
|
87 |
self.ema_model = EMA(model, include_online_model=False, **ema_kwargs)
|
88 |
-
|
89 |
self.ema_model.to(self.accelerator.device)
|
90 |
|
91 |
self.epochs = epochs
|
@@ -176,6 +187,14 @@ class Trainer:
|
|
176 |
return step
|
177 |
|
178 |
def train(self, train_dataset: Dataset, num_workers=16, resumable_with_seed: int = None):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
179 |
if exists(resumable_with_seed):
|
180 |
generator = torch.Generator()
|
181 |
generator.manual_seed(resumable_with_seed)
|
@@ -286,12 +305,31 @@ class Trainer:
|
|
286 |
|
287 |
if self.accelerator.is_local_main_process:
|
288 |
self.accelerator.log({"loss": loss.item(), "lr": self.scheduler.get_last_lr()[0]}, step=global_step)
|
|
|
|
|
|
|
289 |
|
290 |
progress_bar.set_postfix(step=str(global_step), loss=loss.item())
|
291 |
|
292 |
if global_step % (self.save_per_updates * self.grad_accumulation_steps) == 0:
|
293 |
self.save_checkpoint(global_step)
|
294 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
295 |
if global_step % self.last_per_steps == 0:
|
296 |
self.save_checkpoint(global_step, last=True)
|
297 |
|
|
|
6 |
import wandb
|
7 |
|
8 |
import torch
|
9 |
+
import torchaudio
|
10 |
from torch.optim import AdamW
|
11 |
from torch.utils.data import DataLoader, Dataset, SequentialSampler
|
12 |
from torch.optim.lr_scheduler import LinearLR, SequentialLR
|
|
|
40 |
max_grad_norm=1.0,
|
41 |
noise_scheduler: str | None = None,
|
42 |
duration_predictor: torch.nn.Module | None = None,
|
43 |
+
logger: str | None = "wandb", # "wandb" | "tensorboard" | None
|
44 |
wandb_project="test_e2-tts",
|
45 |
wandb_run_name="test_run",
|
46 |
wandb_resume_id: str = None,
|
47 |
+
log_samples: bool = False,
|
48 |
last_per_steps=None,
|
49 |
accelerate_kwargs: dict = dict(),
|
50 |
ema_kwargs: dict = dict(),
|
|
|
52 |
):
|
53 |
ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
|
54 |
|
55 |
+
if logger == "wandb" and not wandb.api.api_key:
|
56 |
+
logger = None
|
57 |
print(f"Using logger: {logger}")
|
58 |
+
self.log_samples = log_samples
|
59 |
|
60 |
self.accelerator = Accelerator(
|
61 |
+
log_with=logger if logger == "wandb" else None,
|
62 |
kwargs_handlers=[ddp_kwargs],
|
63 |
gradient_accumulation_steps=grad_accumulation_steps,
|
64 |
**accelerate_kwargs,
|
65 |
)
|
66 |
|
67 |
+
self.logger = logger
|
68 |
+
if self.logger == "wandb":
|
69 |
if exists(wandb_resume_id):
|
70 |
init_kwargs = {"wandb": {"resume": "allow", "name": wandb_run_name, "id": wandb_resume_id}}
|
71 |
else:
|
72 |
init_kwargs = {"wandb": {"resume": "allow", "name": wandb_run_name}}
|
73 |
+
|
74 |
self.accelerator.init_trackers(
|
75 |
project_name=wandb_project,
|
76 |
init_kwargs=init_kwargs,
|
|
|
88 |
},
|
89 |
)
|
90 |
|
91 |
+
elif self.logger == "tensorboard":
|
92 |
+
from torch.utils.tensorboard import SummaryWriter
|
93 |
+
|
94 |
+
self.writer = SummaryWriter(log_dir=f"runs/{wandb_run_name}")
|
95 |
+
|
96 |
self.model = model
|
97 |
|
98 |
if self.is_main:
|
99 |
self.ema_model = EMA(model, include_online_model=False, **ema_kwargs)
|
|
|
100 |
self.ema_model.to(self.accelerator.device)
|
101 |
|
102 |
self.epochs = epochs
|
|
|
187 |
return step
|
188 |
|
189 |
def train(self, train_dataset: Dataset, num_workers=16, resumable_with_seed: int = None):
|
190 |
+
if self.log_samples:
|
191 |
+
from f5_tts.infer.utils_infer import load_vocoder, nfe_step, cfg_strength, sway_sampling_coef
|
192 |
+
|
193 |
+
vocoder = load_vocoder()
|
194 |
+
target_sample_rate = self.accelerator.unwrap_model(self.model).mel_spec.mel_stft.sample_rate
|
195 |
+
log_samples_path = f"{self.checkpoint_path}/samples"
|
196 |
+
os.makedirs(log_samples_path, exist_ok=True)
|
197 |
+
|
198 |
if exists(resumable_with_seed):
|
199 |
generator = torch.Generator()
|
200 |
generator.manual_seed(resumable_with_seed)
|
|
|
305 |
|
306 |
if self.accelerator.is_local_main_process:
|
307 |
self.accelerator.log({"loss": loss.item(), "lr": self.scheduler.get_last_lr()[0]}, step=global_step)
|
308 |
+
if self.logger == "tensorboard":
|
309 |
+
self.writer.add_scalar("loss", loss.item(), global_step)
|
310 |
+
self.writer.add_scalar("lr", self.scheduler.get_last_lr()[0], global_step)
|
311 |
|
312 |
progress_bar.set_postfix(step=str(global_step), loss=loss.item())
|
313 |
|
314 |
if global_step % (self.save_per_updates * self.grad_accumulation_steps) == 0:
|
315 |
self.save_checkpoint(global_step)
|
316 |
|
317 |
+
if self.log_samples and self.accelerator.is_local_main_process:
|
318 |
+
ref_audio, ref_audio_len = vocoder.decode(batch["mel"][0].unsqueeze(0).cpu()), mel_lengths[0]
|
319 |
+
torchaudio.save(f"{log_samples_path}/step_{global_step}_ref.wav", ref_audio, target_sample_rate)
|
320 |
+
with torch.inference_mode():
|
321 |
+
generated, _ = self.accelerator.unwrap_model(self.model).sample(
|
322 |
+
cond=mel_spec[0][:ref_audio_len].unsqueeze(0),
|
323 |
+
text=[text_inputs[0] + [" "] + text_inputs[0]],
|
324 |
+
duration=ref_audio_len * 2,
|
325 |
+
steps=nfe_step,
|
326 |
+
cfg_strength=cfg_strength,
|
327 |
+
sway_sampling_coef=sway_sampling_coef,
|
328 |
+
)
|
329 |
+
generated = generated.to(torch.float32)
|
330 |
+
gen_audio = vocoder.decode(generated[:, ref_audio_len:, :].permute(0, 2, 1).cpu())
|
331 |
+
torchaudio.save(f"{log_samples_path}/step_{global_step}_gen.wav", gen_audio, target_sample_rate)
|
332 |
+
|
333 |
if global_step % self.last_per_steps == 0:
|
334 |
self.save_checkpoint(global_step, last=True)
|
335 |
|
src/f5_tts/train/finetune_cli.py
CHANGED
@@ -56,6 +56,14 @@ def parse_args():
|
|
56 |
help="Path to custom tokenizer vocab file (only used if tokenizer = 'custom')",
|
57 |
)
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
return parser.parse_args()
|
60 |
|
61 |
|
@@ -64,6 +72,7 @@ def parse_args():
|
|
64 |
|
65 |
def main():
|
66 |
args = parse_args()
|
|
|
67 |
checkpoint_path = str(files("f5_tts").joinpath(f"../../ckpts/{args.dataset_name}"))
|
68 |
|
69 |
# Model parameters based on experiment name
|
@@ -132,9 +141,11 @@ def main():
|
|
132 |
max_samples=args.max_samples,
|
133 |
grad_accumulation_steps=args.grad_accumulation_steps,
|
134 |
max_grad_norm=args.max_grad_norm,
|
|
|
135 |
wandb_project=args.dataset_name,
|
136 |
wandb_run_name=args.exp_name,
|
137 |
wandb_resume_id=wandb_resume_id,
|
|
|
138 |
last_per_steps=args.last_per_steps,
|
139 |
)
|
140 |
|
|
|
56 |
help="Path to custom tokenizer vocab file (only used if tokenizer = 'custom')",
|
57 |
)
|
58 |
|
59 |
+
parser.add_argument(
|
60 |
+
"--log_samples",
|
61 |
+
type=bool,
|
62 |
+
default=False,
|
63 |
+
help="Log inferenced samples per ckpt save steps",
|
64 |
+
)
|
65 |
+
parser.add_argument("--logger", type=str, default=None, choices=["wandb", "tensorboard"], help="logger")
|
66 |
+
|
67 |
return parser.parse_args()
|
68 |
|
69 |
|
|
|
72 |
|
73 |
def main():
|
74 |
args = parse_args()
|
75 |
+
|
76 |
checkpoint_path = str(files("f5_tts").joinpath(f"../../ckpts/{args.dataset_name}"))
|
77 |
|
78 |
# Model parameters based on experiment name
|
|
|
141 |
max_samples=args.max_samples,
|
142 |
grad_accumulation_steps=args.grad_accumulation_steps,
|
143 |
max_grad_norm=args.max_grad_norm,
|
144 |
+
logger=args.logger,
|
145 |
wandb_project=args.dataset_name,
|
146 |
wandb_run_name=args.exp_name,
|
147 |
wandb_resume_id=wandb_resume_id,
|
148 |
+
log_samples=args.log_samples,
|
149 |
last_per_steps=args.last_per_steps,
|
150 |
)
|
151 |
|
src/f5_tts/train/finetune_gradio.py
CHANGED
@@ -69,6 +69,7 @@ def save_settings(
|
|
69 |
tokenizer_type,
|
70 |
tokenizer_file,
|
71 |
mixed_precision,
|
|
|
72 |
):
|
73 |
path_project = os.path.join(path_project_ckpts, project_name)
|
74 |
os.makedirs(path_project, exist_ok=True)
|
@@ -91,6 +92,7 @@ def save_settings(
|
|
91 |
"tokenizer_type": tokenizer_type,
|
92 |
"tokenizer_file": tokenizer_file,
|
93 |
"mixed_precision": mixed_precision,
|
|
|
94 |
}
|
95 |
with open(file_setting, "w") as f:
|
96 |
json.dump(settings, f, indent=4)
|
@@ -121,6 +123,7 @@ def load_settings(project_name):
|
|
121 |
"tokenizer_type": "pinyin",
|
122 |
"tokenizer_file": "",
|
123 |
"mixed_precision": "none",
|
|
|
124 |
}
|
125 |
return (
|
126 |
settings["exp_name"],
|
@@ -139,6 +142,7 @@ def load_settings(project_name):
|
|
139 |
settings["tokenizer_type"],
|
140 |
settings["tokenizer_file"],
|
141 |
settings["mixed_precision"],
|
|
|
142 |
)
|
143 |
|
144 |
with open(file_setting, "r") as f:
|
@@ -160,6 +164,7 @@ def load_settings(project_name):
|
|
160 |
settings["tokenizer_type"],
|
161 |
settings["tokenizer_file"],
|
162 |
settings["mixed_precision"],
|
|
|
163 |
)
|
164 |
|
165 |
|
@@ -374,6 +379,7 @@ def start_training(
|
|
374 |
tokenizer_file="",
|
375 |
mixed_precision="fp16",
|
376 |
stream=False,
|
|
|
377 |
):
|
378 |
global training_process, tts_api, stop_signal
|
379 |
|
@@ -447,6 +453,8 @@ def start_training(
|
|
447 |
|
448 |
cmd += f" --tokenizer {tokenizer_type} "
|
449 |
|
|
|
|
|
450 |
print(cmd)
|
451 |
|
452 |
save_settings(
|
@@ -467,6 +475,7 @@ def start_training(
|
|
467 |
tokenizer_type,
|
468 |
tokenizer_file,
|
469 |
mixed_precision,
|
|
|
470 |
)
|
471 |
|
472 |
try:
|
@@ -1223,6 +1232,27 @@ def get_checkpoints_project(project_name, is_gradio=True):
|
|
1223 |
return files_checkpoints, selelect_checkpoint
|
1224 |
|
1225 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1226 |
def get_gpu_stats():
|
1227 |
gpu_stats = ""
|
1228 |
|
@@ -1290,6 +1320,17 @@ def get_combined_stats():
|
|
1290 |
return combined_stats
|
1291 |
|
1292 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1293 |
with gr.Blocks() as app:
|
1294 |
gr.Markdown(
|
1295 |
"""
|
@@ -1470,6 +1511,7 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1470 |
|
1471 |
with gr.Row():
|
1472 |
mixed_precision = gr.Radio(label="mixed_precision", choices=["none", "fp16", "fpb16"], value="none")
|
|
|
1473 |
start_button = gr.Button("Start Training")
|
1474 |
stop_button = gr.Button("Stop Training", interactive=False)
|
1475 |
|
@@ -1491,6 +1533,7 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1491 |
tokenizer_typev,
|
1492 |
tokenizer_filev,
|
1493 |
mixed_precisionv,
|
|
|
1494 |
) = load_settings(projects_selelect)
|
1495 |
exp_name.value = exp_namev
|
1496 |
learning_rate.value = learning_ratev
|
@@ -1508,9 +1551,43 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1508 |
tokenizer_type.value = tokenizer_typev
|
1509 |
tokenizer_file.value = tokenizer_filev
|
1510 |
mixed_precision.value = mixed_precisionv
|
|
|
1511 |
|
1512 |
ch_stream = gr.Checkbox(label="stream output experiment.", value=True)
|
1513 |
txt_info_train = gr.Text(label="info", value="")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1514 |
start_button.click(
|
1515 |
fn=start_training,
|
1516 |
inputs=[
|
@@ -1532,6 +1609,7 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1532 |
tokenizer_file,
|
1533 |
mixed_precision,
|
1534 |
ch_stream,
|
|
|
1535 |
],
|
1536 |
outputs=[txt_info_train, start_button, stop_button],
|
1537 |
)
|
@@ -1583,6 +1661,7 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
|
|
1583 |
tokenizer_type,
|
1584 |
tokenizer_file,
|
1585 |
mixed_precision,
|
|
|
1586 |
]
|
1587 |
|
1588 |
return output_components
|
|
|
69 |
tokenizer_type,
|
70 |
tokenizer_file,
|
71 |
mixed_precision,
|
72 |
+
logger,
|
73 |
):
|
74 |
path_project = os.path.join(path_project_ckpts, project_name)
|
75 |
os.makedirs(path_project, exist_ok=True)
|
|
|
92 |
"tokenizer_type": tokenizer_type,
|
93 |
"tokenizer_file": tokenizer_file,
|
94 |
"mixed_precision": mixed_precision,
|
95 |
+
"logger": logger,
|
96 |
}
|
97 |
with open(file_setting, "w") as f:
|
98 |
json.dump(settings, f, indent=4)
|
|
|
123 |
"tokenizer_type": "pinyin",
|
124 |
"tokenizer_file": "",
|
125 |
"mixed_precision": "none",
|
126 |
+
"logger": "wandb",
|
127 |
}
|
128 |
return (
|
129 |
settings["exp_name"],
|
|
|
142 |
settings["tokenizer_type"],
|
143 |
settings["tokenizer_file"],
|
144 |
settings["mixed_precision"],
|
145 |
+
settings["logger"],
|
146 |
)
|
147 |
|
148 |
with open(file_setting, "r") as f:
|
|
|
164 |
settings["tokenizer_type"],
|
165 |
settings["tokenizer_file"],
|
166 |
settings["mixed_precision"],
|
167 |
+
settings["logger"],
|
168 |
)
|
169 |
|
170 |
|
|
|
379 |
tokenizer_file="",
|
380 |
mixed_precision="fp16",
|
381 |
stream=False,
|
382 |
+
logger="wandb",
|
383 |
):
|
384 |
global training_process, tts_api, stop_signal
|
385 |
|
|
|
453 |
|
454 |
cmd += f" --tokenizer {tokenizer_type} "
|
455 |
|
456 |
+
cmd += f" --log_samples True --logger {logger} "
|
457 |
+
|
458 |
print(cmd)
|
459 |
|
460 |
save_settings(
|
|
|
475 |
tokenizer_type,
|
476 |
tokenizer_file,
|
477 |
mixed_precision,
|
478 |
+
logger,
|
479 |
)
|
480 |
|
481 |
try:
|
|
|
1232 |
return files_checkpoints, selelect_checkpoint
|
1233 |
|
1234 |
|
1235 |
+
def get_audio_project(project_name, is_gradio=True):
|
1236 |
+
if project_name is None:
|
1237 |
+
return [], ""
|
1238 |
+
project_name = project_name.replace("_pinyin", "").replace("_char", "")
|
1239 |
+
|
1240 |
+
if os.path.isdir(path_project_ckpts):
|
1241 |
+
files_audios = glob(os.path.join(path_project_ckpts, project_name, "samples", "*.wav"))
|
1242 |
+
files_audios = sorted(files_audios, key=lambda x: int(os.path.basename(x).split("_")[1].split(".")[0]))
|
1243 |
+
|
1244 |
+
files_audios = [item.replace("_gen.wav", "") for item in files_audios if item.endswith("_gen.wav")]
|
1245 |
+
else:
|
1246 |
+
files_audios = []
|
1247 |
+
|
1248 |
+
selelect_checkpoint = None if not files_audios else files_audios[0]
|
1249 |
+
|
1250 |
+
if is_gradio:
|
1251 |
+
return gr.update(choices=files_audios, value=selelect_checkpoint)
|
1252 |
+
|
1253 |
+
return files_audios, selelect_checkpoint
|
1254 |
+
|
1255 |
+
|
1256 |
def get_gpu_stats():
|
1257 |
gpu_stats = ""
|
1258 |
|
|
|
1320 |
return combined_stats
|
1321 |
|
1322 |
|
1323 |
+
def get_audio_select(file_sample):
|
1324 |
+
select_audio_ref = file_sample
|
1325 |
+
select_audio_gen = file_sample
|
1326 |
+
|
1327 |
+
if file_sample is not None:
|
1328 |
+
select_audio_ref += "_ref.wav"
|
1329 |
+
select_audio_gen += "_gen.wav"
|
1330 |
+
|
1331 |
+
return select_audio_ref, select_audio_gen
|
1332 |
+
|
1333 |
+
|
1334 |
with gr.Blocks() as app:
|
1335 |
gr.Markdown(
|
1336 |
"""
|
|
|
1511 |
|
1512 |
with gr.Row():
|
1513 |
mixed_precision = gr.Radio(label="mixed_precision", choices=["none", "fp16", "fpb16"], value="none")
|
1514 |
+
cd_logger = gr.Radio(label="logger", choices=["wandb", "tensorboard"], value="wandb")
|
1515 |
start_button = gr.Button("Start Training")
|
1516 |
stop_button = gr.Button("Stop Training", interactive=False)
|
1517 |
|
|
|
1533 |
tokenizer_typev,
|
1534 |
tokenizer_filev,
|
1535 |
mixed_precisionv,
|
1536 |
+
cd_loggerv,
|
1537 |
) = load_settings(projects_selelect)
|
1538 |
exp_name.value = exp_namev
|
1539 |
learning_rate.value = learning_ratev
|
|
|
1551 |
tokenizer_type.value = tokenizer_typev
|
1552 |
tokenizer_file.value = tokenizer_filev
|
1553 |
mixed_precision.value = mixed_precisionv
|
1554 |
+
cd_logger.value = cd_loggerv
|
1555 |
|
1556 |
ch_stream = gr.Checkbox(label="stream output experiment.", value=True)
|
1557 |
txt_info_train = gr.Text(label="info", value="")
|
1558 |
+
|
1559 |
+
list_audios, select_audio = get_audio_project(projects_selelect, False)
|
1560 |
+
|
1561 |
+
select_audio_ref = select_audio
|
1562 |
+
select_audio_gen = select_audio
|
1563 |
+
|
1564 |
+
if select_audio is not None:
|
1565 |
+
select_audio_ref += "_ref.wav"
|
1566 |
+
select_audio_gen += "_gen.wav"
|
1567 |
+
|
1568 |
+
with gr.Row():
|
1569 |
+
ch_list_audio = gr.Dropdown(
|
1570 |
+
choices=list_audios,
|
1571 |
+
value=select_audio,
|
1572 |
+
label="audios",
|
1573 |
+
allow_custom_value=True,
|
1574 |
+
scale=6,
|
1575 |
+
interactive=True,
|
1576 |
+
)
|
1577 |
+
bt_stream_audio = gr.Button("refresh", scale=1)
|
1578 |
+
bt_stream_audio.click(fn=get_audio_project, inputs=[cm_project], outputs=[ch_list_audio])
|
1579 |
+
cm_project.change(fn=get_audio_project, inputs=[cm_project], outputs=[ch_list_audio])
|
1580 |
+
|
1581 |
+
with gr.Row():
|
1582 |
+
audio_ref_stream = gr.Audio(label="original", type="filepath", value=select_audio_ref)
|
1583 |
+
audio_gen_stream = gr.Audio(label="generate", type="filepath", value=select_audio_gen)
|
1584 |
+
|
1585 |
+
ch_list_audio.change(
|
1586 |
+
fn=get_audio_select,
|
1587 |
+
inputs=[ch_list_audio],
|
1588 |
+
outputs=[audio_ref_stream, audio_gen_stream],
|
1589 |
+
)
|
1590 |
+
|
1591 |
start_button.click(
|
1592 |
fn=start_training,
|
1593 |
inputs=[
|
|
|
1609 |
tokenizer_file,
|
1610 |
mixed_precision,
|
1611 |
ch_stream,
|
1612 |
+
cd_logger,
|
1613 |
],
|
1614 |
outputs=[txt_info_train, start_button, stop_button],
|
1615 |
)
|
|
|
1661 |
tokenizer_type,
|
1662 |
tokenizer_file,
|
1663 |
mixed_precision,
|
1664 |
+
cd_logger,
|
1665 |
]
|
1666 |
|
1667 |
return output_components
|
src/f5_tts/train/train.py
CHANGED
@@ -83,6 +83,7 @@ def main():
|
|
83 |
wandb_run_name=exp_name,
|
84 |
wandb_resume_id=wandb_resume_id,
|
85 |
last_per_steps=last_per_steps,
|
|
|
86 |
)
|
87 |
|
88 |
train_dataset = load_dataset(dataset_name, tokenizer, mel_spec_kwargs=mel_spec_kwargs)
|
|
|
83 |
wandb_run_name=exp_name,
|
84 |
wandb_resume_id=wandb_resume_id,
|
85 |
last_per_steps=last_per_steps,
|
86 |
+
log_samples=True,
|
87 |
)
|
88 |
|
89 |
train_dataset = load_dataset(dataset_name, tokenizer, mel_spec_kwargs=mel_spec_kwargs)
|