Spaces:
Sleeping
Sleeping
Commit
·
90cacdf
1
Parent(s):
e4c0874
Fix folder structure
Browse files- .gitignore +0 -1
- datasets.py → remfx/datasets.py +0 -0
- remfx/models.py +192 -0
- remfx/utils.py +71 -0
- download_egfx.sh → scripts/download_egfx.sh +0 -0
- train.py → scripts/train.py +0 -0
.gitignore
CHANGED
|
@@ -6,7 +6,6 @@ data/
|
|
| 6 |
.DS_Store
|
| 7 |
__pycache__/
|
| 8 |
lightning_logs/
|
| 9 |
-
RemFX/
|
| 10 |
outputs/
|
| 11 |
logs/
|
| 12 |
.vscode/
|
|
|
|
| 6 |
.DS_Store
|
| 7 |
__pycache__/
|
| 8 |
lightning_logs/
|
|
|
|
| 9 |
outputs/
|
| 10 |
logs/
|
| 11 |
.vscode/
|
datasets.py → remfx/datasets.py
RENAMED
|
File without changes
|
remfx/models.py
ADDED
|
@@ -0,0 +1,192 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from torch import Tensor, nn
|
| 3 |
+
import pytorch_lightning as pl
|
| 4 |
+
from einops import rearrange
|
| 5 |
+
import wandb
|
| 6 |
+
from audio_diffusion_pytorch import DiffusionModel
|
| 7 |
+
import auraloss
|
| 8 |
+
|
| 9 |
+
from umx.openunmix.model import OpenUnmix, Separator
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
class RemFXModel(pl.LightningModule):
|
| 13 |
+
def __init__(
|
| 14 |
+
self,
|
| 15 |
+
lr: float,
|
| 16 |
+
lr_beta1: float,
|
| 17 |
+
lr_beta2: float,
|
| 18 |
+
lr_eps: float,
|
| 19 |
+
lr_weight_decay: float,
|
| 20 |
+
sample_rate: float,
|
| 21 |
+
network: nn.Module,
|
| 22 |
+
):
|
| 23 |
+
super().__init__()
|
| 24 |
+
self.lr = lr
|
| 25 |
+
self.lr_beta1 = lr_beta1
|
| 26 |
+
self.lr_beta2 = lr_beta2
|
| 27 |
+
self.lr_eps = lr_eps
|
| 28 |
+
self.lr_weight_decay = lr_weight_decay
|
| 29 |
+
self.sample_rate = sample_rate
|
| 30 |
+
self.model = network
|
| 31 |
+
|
| 32 |
+
@property
|
| 33 |
+
def device(self):
|
| 34 |
+
return next(self.model.parameters()).device
|
| 35 |
+
|
| 36 |
+
def configure_optimizers(self):
|
| 37 |
+
optimizer = torch.optim.AdamW(
|
| 38 |
+
list(self.model.parameters()),
|
| 39 |
+
lr=self.lr,
|
| 40 |
+
betas=(self.lr_beta1, self.lr_beta2),
|
| 41 |
+
eps=self.lr_eps,
|
| 42 |
+
weight_decay=self.lr_weight_decay,
|
| 43 |
+
)
|
| 44 |
+
return optimizer
|
| 45 |
+
|
| 46 |
+
def training_step(self, batch, batch_idx):
|
| 47 |
+
loss = self.common_step(batch, batch_idx, mode="train")
|
| 48 |
+
return loss
|
| 49 |
+
|
| 50 |
+
def validation_step(self, batch, batch_idx):
|
| 51 |
+
loss = self.common_step(batch, batch_idx, mode="valid")
|
| 52 |
+
|
| 53 |
+
def common_step(self, batch, batch_idx, mode: str = "train"):
|
| 54 |
+
loss = self.model(batch)
|
| 55 |
+
self.log(f"{mode}_loss", loss)
|
| 56 |
+
return loss
|
| 57 |
+
|
| 58 |
+
def on_validation_epoch_start(self):
|
| 59 |
+
self.log_next = True
|
| 60 |
+
|
| 61 |
+
def on_validation_batch_start(self, batch, batch_idx, dataloader_idx):
|
| 62 |
+
if self.log_next:
|
| 63 |
+
x, target, label = batch
|
| 64 |
+
y = self.model.sample(x)
|
| 65 |
+
log_wandb_audio_batch(
|
| 66 |
+
logger=self.logger,
|
| 67 |
+
id="sample",
|
| 68 |
+
samples=x.cpu(),
|
| 69 |
+
sampling_rate=self.sample_rate,
|
| 70 |
+
caption=f"Epoch {self.current_epoch}",
|
| 71 |
+
)
|
| 72 |
+
log_wandb_audio_batch(
|
| 73 |
+
logger=self.logger,
|
| 74 |
+
id="prediction",
|
| 75 |
+
samples=y.cpu(),
|
| 76 |
+
sampling_rate=self.sample_rate,
|
| 77 |
+
caption=f"Epoch {self.current_epoch}",
|
| 78 |
+
)
|
| 79 |
+
log_wandb_audio_batch(
|
| 80 |
+
logger=self.logger,
|
| 81 |
+
id="target",
|
| 82 |
+
samples=target.cpu(),
|
| 83 |
+
sampling_rate=self.sample_rate,
|
| 84 |
+
caption=f"Epoch {self.current_epoch}",
|
| 85 |
+
)
|
| 86 |
+
self.log_next = False
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
class OpenUnmixModel(torch.nn.Module):
|
| 90 |
+
def __init__(
|
| 91 |
+
self,
|
| 92 |
+
n_fft: int = 2048,
|
| 93 |
+
hop_length: int = 512,
|
| 94 |
+
n_channels: int = 1,
|
| 95 |
+
alpha: float = 0.3,
|
| 96 |
+
sample_rate: int = 22050,
|
| 97 |
+
):
|
| 98 |
+
super().__init__()
|
| 99 |
+
self.n_channels = n_channels
|
| 100 |
+
self.n_fft = n_fft
|
| 101 |
+
self.hop_length = hop_length
|
| 102 |
+
self.alpha = alpha
|
| 103 |
+
window = torch.hann_window(n_fft)
|
| 104 |
+
self.register_buffer("window", window)
|
| 105 |
+
|
| 106 |
+
self.num_bins = self.n_fft // 2 + 1
|
| 107 |
+
self.sample_rate = sample_rate
|
| 108 |
+
self.model = OpenUnmix(
|
| 109 |
+
nb_channels=self.n_channels,
|
| 110 |
+
nb_bins=self.num_bins,
|
| 111 |
+
)
|
| 112 |
+
self.separator = Separator(
|
| 113 |
+
target_models={"other": self.model},
|
| 114 |
+
nb_channels=self.n_channels,
|
| 115 |
+
sample_rate=self.sample_rate,
|
| 116 |
+
n_fft=self.n_fft,
|
| 117 |
+
n_hop=self.hop_length,
|
| 118 |
+
)
|
| 119 |
+
self.loss_fn = auraloss.freq.MultiResolutionSTFTLoss(
|
| 120 |
+
n_bins=self.num_bins, sample_rate=self.sample_rate
|
| 121 |
+
)
|
| 122 |
+
|
| 123 |
+
def forward(self, batch):
|
| 124 |
+
x, target, label = batch
|
| 125 |
+
X = spectrogram(x, self.window, self.n_fft, self.hop_length, self.alpha)
|
| 126 |
+
Y = self.model(X)
|
| 127 |
+
sep_out = self.separator(x).squeeze(1)
|
| 128 |
+
loss = self.loss_fn(sep_out, target)
|
| 129 |
+
|
| 130 |
+
return loss
|
| 131 |
+
|
| 132 |
+
def sample(self, x: Tensor) -> Tensor:
|
| 133 |
+
return self.separator(x).squeeze(1)
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
class DiffusionGenerationModel(nn.Module):
|
| 137 |
+
def __init__(self, n_channels: int = 1):
|
| 138 |
+
super().__init__()
|
| 139 |
+
self.model = DiffusionModel(in_channels=n_channels)
|
| 140 |
+
|
| 141 |
+
def forward(self, batch):
|
| 142 |
+
x, target, label = batch
|
| 143 |
+
return self.model(x)
|
| 144 |
+
|
| 145 |
+
def sample(self, x: Tensor, num_steps: int = 10) -> Tensor:
|
| 146 |
+
noise = torch.randn(x.shape).to(x)
|
| 147 |
+
return self.model.sample(noise, num_steps=num_steps)
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
def log_wandb_audio_batch(
|
| 151 |
+
logger: pl.loggers.WandbLogger,
|
| 152 |
+
id: str,
|
| 153 |
+
samples: Tensor,
|
| 154 |
+
sampling_rate: int,
|
| 155 |
+
caption: str = "",
|
| 156 |
+
):
|
| 157 |
+
num_items = samples.shape[0]
|
| 158 |
+
samples = rearrange(samples, "b c t -> b t c")
|
| 159 |
+
for idx in range(num_items):
|
| 160 |
+
logger.experiment.log(
|
| 161 |
+
{
|
| 162 |
+
f"{id}_{idx}": wandb.Audio(
|
| 163 |
+
samples[idx].cpu().numpy(),
|
| 164 |
+
caption=caption,
|
| 165 |
+
sample_rate=sampling_rate,
|
| 166 |
+
)
|
| 167 |
+
}
|
| 168 |
+
)
|
| 169 |
+
|
| 170 |
+
|
| 171 |
+
def spectrogram(
|
| 172 |
+
x: torch.Tensor,
|
| 173 |
+
window: torch.Tensor,
|
| 174 |
+
n_fft: int,
|
| 175 |
+
hop_length: int,
|
| 176 |
+
alpha: float,
|
| 177 |
+
) -> torch.Tensor:
|
| 178 |
+
bs, chs, samp = x.size()
|
| 179 |
+
x = x.view(bs * chs, -1) # move channels onto batch dim
|
| 180 |
+
|
| 181 |
+
X = torch.stft(
|
| 182 |
+
x,
|
| 183 |
+
n_fft=n_fft,
|
| 184 |
+
hop_length=hop_length,
|
| 185 |
+
window=window,
|
| 186 |
+
return_complex=True,
|
| 187 |
+
)
|
| 188 |
+
|
| 189 |
+
# move channels back
|
| 190 |
+
X = X.view(bs, chs, X.shape[-2], X.shape[-1])
|
| 191 |
+
|
| 192 |
+
return torch.pow(X.abs() + 1e-8, alpha)
|
remfx/utils.py
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import logging
|
| 2 |
+
from typing import List
|
| 3 |
+
import pytorch_lightning as pl
|
| 4 |
+
from omegaconf import DictConfig
|
| 5 |
+
from pytorch_lightning.utilities import rank_zero_only
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
def get_logger(name=__name__) -> logging.Logger:
|
| 9 |
+
"""Initializes multi-GPU-friendly python command line logger."""
|
| 10 |
+
|
| 11 |
+
logger = logging.getLogger(name)
|
| 12 |
+
|
| 13 |
+
# this ensures all logging levels get marked with the rank zero decorator
|
| 14 |
+
# otherwise logs would get multiplied for each GPU process in multi-GPU setup
|
| 15 |
+
for level in (
|
| 16 |
+
"debug",
|
| 17 |
+
"info",
|
| 18 |
+
"warning",
|
| 19 |
+
"error",
|
| 20 |
+
"exception",
|
| 21 |
+
"fatal",
|
| 22 |
+
"critical",
|
| 23 |
+
):
|
| 24 |
+
setattr(logger, level, rank_zero_only(getattr(logger, level)))
|
| 25 |
+
|
| 26 |
+
return logger
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
log = get_logger(__name__)
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
@rank_zero_only
|
| 33 |
+
def log_hyperparameters(
|
| 34 |
+
config: DictConfig,
|
| 35 |
+
model: pl.LightningModule,
|
| 36 |
+
datamodule: pl.LightningDataModule,
|
| 37 |
+
trainer: pl.Trainer,
|
| 38 |
+
callbacks: List[pl.Callback],
|
| 39 |
+
logger: pl.loggers.logger.Logger,
|
| 40 |
+
) -> None:
|
| 41 |
+
"""Controls which config parts are saved by Lightning loggers.
|
| 42 |
+
Additionaly saves:
|
| 43 |
+
- number of model parameters
|
| 44 |
+
"""
|
| 45 |
+
|
| 46 |
+
if not trainer.logger:
|
| 47 |
+
return
|
| 48 |
+
|
| 49 |
+
hparams = {}
|
| 50 |
+
|
| 51 |
+
# choose which parts of hydra config will be saved to loggers
|
| 52 |
+
hparams["model"] = config["model"]
|
| 53 |
+
|
| 54 |
+
# save number of model parameters
|
| 55 |
+
hparams["model/params/total"] = sum(p.numel() for p in model.parameters())
|
| 56 |
+
hparams["model/params/trainable"] = sum(
|
| 57 |
+
p.numel() for p in model.parameters() if p.requires_grad
|
| 58 |
+
)
|
| 59 |
+
hparams["model/params/non_trainable"] = sum(
|
| 60 |
+
p.numel() for p in model.parameters() if not p.requires_grad
|
| 61 |
+
)
|
| 62 |
+
|
| 63 |
+
hparams["datamodule"] = config["datamodule"]
|
| 64 |
+
hparams["trainer"] = config["trainer"]
|
| 65 |
+
|
| 66 |
+
if "seed" in config:
|
| 67 |
+
hparams["seed"] = config["seed"]
|
| 68 |
+
if "callbacks" in config:
|
| 69 |
+
hparams["callbacks"] = config["callbacks"]
|
| 70 |
+
|
| 71 |
+
logger.experiment.config.update(hparams)
|
download_egfx.sh → scripts/download_egfx.sh
RENAMED
|
File without changes
|
train.py → scripts/train.py
RENAMED
|
File without changes
|