Spaces:
No application file
No application file
File size: 7,704 Bytes
5f84dff |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 |
from argparse import ArgumentParser
import matplotlib.pyplot as plt
import pytorch_lightning as pl
import torch
import wandb
from loguru import logger
from mmengine import Config
from mmengine.optim import OPTIMIZERS
from pytorch_lightning.loggers import TensorBoardLogger, WandbLogger
from torch.utils.data import DataLoader
from fish_diffusion.archs.diffsinger import DiffSinger
from fish_diffusion.datasets import DATASETS
from fish_diffusion.datasets.repeat import RepeatDataset
from fish_diffusion.utils.scheduler import LR_SCHEUDLERS
from fish_diffusion.utils.viz import viz_synth_sample
from fish_diffusion.vocoders import VOCODERS
class FishDiffusion(pl.LightningModule):
def __init__(self, config):
super().__init__()
self.save_hyperparameters()
self.model = DiffSinger(config.model)
self.config = config
# 音频编码器, 将梅尔谱转换为音频
self.vocoder = VOCODERS.build(config.model.vocoder)
self.vocoder.freeze()
def configure_optimizers(self):
self.config.optimizer.params = self.parameters()
optimizer = OPTIMIZERS.build(self.config.optimizer)
self.config.scheduler.optimizer = optimizer
scheduler = LR_SCHEUDLERS.build(self.config.scheduler)
return [optimizer], dict(scheduler=scheduler, interval="step")
def _step(self, batch, batch_idx, mode):
assert batch["pitches"].shape[1] == batch["mels"].shape[1]
pitches = batch["pitches"].clone()
batch_size = batch["speakers"].shape[0]
output = self.model(
speakers=batch["speakers"],
contents=batch["contents"],
src_lens=batch["content_lens"],
max_src_len=batch["max_content_len"],
mels=batch["mels"],
mel_lens=batch["mel_lens"],
max_mel_len=batch["max_mel_len"],
pitches=batch["pitches"],
)
self.log(f"{mode}_loss", output["loss"], batch_size=batch_size, sync_dist=True)
if mode != "valid":
return output["loss"]
x = self.model.diffusion(output["features"])
for idx, (gt_mel, gt_pitch, predict_mel, predict_mel_len) in enumerate(
zip(batch["mels"], pitches, x, batch["mel_lens"])
):
image_mels, wav_reconstruction, wav_prediction = viz_synth_sample(
gt_mel=gt_mel,
gt_pitch=gt_pitch,
predict_mel=predict_mel,
predict_mel_len=predict_mel_len,
vocoder=self.vocoder,
return_image=False,
)
wav_reconstruction = wav_reconstruction.to(torch.float32).cpu().numpy()
wav_prediction = wav_prediction.to(torch.float32).cpu().numpy()
# WanDB logger
if isinstance(self.logger, WandbLogger):
self.logger.experiment.log(
{
f"reconstruction_mel": wandb.Image(image_mels, caption="mels"),
f"wavs": [
wandb.Audio(
wav_reconstruction,
sample_rate=44100,
caption=f"reconstruction (gt)",
),
wandb.Audio(
wav_prediction,
sample_rate=44100,
caption=f"prediction",
),
],
},
)
# TensorBoard logger
if isinstance(self.logger, TensorBoardLogger):
self.logger.experiment.add_figure(
f"sample-{idx}/mels",
image_mels,
global_step=self.global_step,
)
self.logger.experiment.add_audio(
f"sample-{idx}/wavs/gt",
wav_reconstruction,
self.global_step,
sample_rate=44100,
)
self.logger.experiment.add_audio(
f"sample-{idx}/wavs/prediction",
wav_prediction,
self.global_step,
sample_rate=44100,
)
if isinstance(image_mels, plt.Figure):
plt.close(image_mels)
return output["loss"]
def training_step(self, batch, batch_idx):
return self._step(batch, batch_idx, mode="train")
def validation_step(self, batch, batch_idx):
return self._step(batch, batch_idx, mode="valid")
if __name__ == "__main__":
pl.seed_everything(42, workers=True)
parser = ArgumentParser()
parser.add_argument("--config", type=str, required=True)
parser.add_argument("--resume", type=str, default=None)
parser.add_argument(
"--tensorboard",
action="store_true",
default=False,
help="Use tensorboard logger, default is wandb.",
)
parser.add_argument("--resume-id", type=str, default=None, help="Wandb run id.")
parser.add_argument("--entity", type=str, default=None, help="Wandb entity.")
parser.add_argument("--name", type=str, default=None, help="Wandb run name.")
parser.add_argument(
"--pretrained", type=str, default=None, help="Pretrained model."
)
parser.add_argument(
"--only-train-speaker-embeddings",
action="store_true",
default=False,
help="Only train speaker embeddings.",
)
args = parser.parse_args()
cfg = Config.fromfile(args.config)
model = FishDiffusion(cfg)
# We only load the state_dict of the model, not the optimizer.
if args.pretrained:
state_dict = torch.load(args.pretrained, map_location="cpu")
if "state_dict" in state_dict:
state_dict = state_dict["state_dict"]
result = model.load_state_dict(state_dict, strict=False)
missing_keys = set(result.missing_keys)
unexpected_keys = set(result.unexpected_keys)
# Make sure incorrect keys are just noise predictor keys.
unexpected_keys = unexpected_keys - set(
i.replace(".naive_noise_predictor.", ".") for i in missing_keys
)
assert len(unexpected_keys) == 0
if args.only_train_speaker_embeddings:
for name, param in model.named_parameters():
if "speaker_encoder" not in name:
param.requires_grad = False
logger.info(
"Only train speaker embeddings, all other parameters are frozen."
)
logger = (
TensorBoardLogger("logs", name=cfg.model.type)
if args.tensorboard
else WandbLogger(
project=cfg.model.type,
save_dir="logs",
log_model=True,
name=args.name,
entity=args.entity,
resume="must" if args.resume_id else False,
id=args.resume_id,
)
)
trainer = pl.Trainer(
logger=logger,
**cfg.trainer,
)
train_dataset = DATASETS.build(cfg.dataset.train)
train_loader = DataLoader(
train_dataset,
collate_fn=train_dataset.collate_fn,
**cfg.dataloader.train,
)
valid_dataset = DATASETS.build(cfg.dataset.valid)
valid_dataset = RepeatDataset(
valid_dataset, repeat=trainer.num_devices, collate_fn=valid_dataset.collate_fn
)
valid_loader = DataLoader(
valid_dataset,
collate_fn=valid_dataset.collate_fn,
**cfg.dataloader.valid,
)
trainer.fit(model, train_loader, valid_loader, ckpt_path=args.resume)
|