Spaces:
Runtime error
Runtime error
Commit
·
9a9a2c9
1
Parent(s):
7d6f241
Update callbacks, debug new models
Browse files- cfg/config.yaml +5 -1
- cfg/model/audio_diffusion.yaml +2 -2
- cfg/model/dcunet.yaml +5 -3
- cfg/model/demucs.yaml +1 -2
- cfg/model/dptnet.yaml +3 -1
- cfg/model/umx.yaml +1 -2
- remfx/callbacks.py +128 -0
- remfx/datasets.py +0 -1
- remfx/dcunet.py +2 -2
- remfx/dptnet.py +1 -2
- remfx/models.py +40 -121
- remfx/utils.py +3 -29
cfg/config.yaml
CHANGED
|
@@ -41,6 +41,11 @@ callbacks:
|
|
| 41 |
learning_rate_monitor:
|
| 42 |
_target_: pytorch_lightning.callbacks.LearningRateMonitor
|
| 43 |
logging_interval: "step"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
datamodule:
|
| 46 |
_target_: remfx.datasets.VocalSetDatamodule
|
|
@@ -116,4 +121,3 @@ trainer:
|
|
| 116 |
devices: 1
|
| 117 |
gradient_clip_val: 10.0
|
| 118 |
max_steps: 50000
|
| 119 |
-
|
|
|
|
| 41 |
learning_rate_monitor:
|
| 42 |
_target_: pytorch_lightning.callbacks.LearningRateMonitor
|
| 43 |
logging_interval: "step"
|
| 44 |
+
audio_logging:
|
| 45 |
+
_target_: remfx.callbacks.AudioCallback
|
| 46 |
+
sample_rate: ${sample_rate}
|
| 47 |
+
metric_logging:
|
| 48 |
+
_target_: remfx.callbacks.MetricCallback
|
| 49 |
|
| 50 |
datamodule:
|
| 51 |
_target_: remfx.datasets.VocalSetDatamodule
|
|
|
|
| 121 |
devices: 1
|
| 122 |
gradient_clip_val: 10.0
|
| 123 |
max_steps: 50000
|
|
|
cfg/model/audio_diffusion.yaml
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
-
_target_: remfx.models.
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
@@ -13,4 +13,4 @@ model:
|
|
| 13 |
datamodule:
|
| 14 |
dataset:
|
| 15 |
effect_types: ["Clean"]
|
| 16 |
-
batch_size: 2
|
|
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
+
_target_: remfx.models.RemFX
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
|
|
| 13 |
datamodule:
|
| 14 |
dataset:
|
| 15 |
effect_types: ["Clean"]
|
| 16 |
+
batch_size: 2
|
cfg/model/dcunet.yaml
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
-
_target_: remfx.models.
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
@@ -9,7 +9,7 @@ model:
|
|
| 9 |
sample_rate: ${sample_rate}
|
| 10 |
network:
|
| 11 |
_target_: remfx.models.DCUNetModel
|
| 12 |
-
spec_dim:
|
| 13 |
hidden_dim: 768
|
| 14 |
filter_len: 512
|
| 15 |
hop_len: 64
|
|
@@ -19,4 +19,6 @@ model:
|
|
| 19 |
refine_layers: 1
|
| 20 |
is_mask: True
|
| 21 |
norm: 'ins'
|
| 22 |
-
act: 'comp'
|
|
|
|
|
|
|
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
+
_target_: remfx.models.RemFX
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
|
|
| 9 |
sample_rate: ${sample_rate}
|
| 10 |
network:
|
| 11 |
_target_: remfx.models.DCUNetModel
|
| 12 |
+
spec_dim: 257
|
| 13 |
hidden_dim: 768
|
| 14 |
filter_len: 512
|
| 15 |
hop_len: 64
|
|
|
|
| 19 |
refine_layers: 1
|
| 20 |
is_mask: True
|
| 21 |
norm: 'ins'
|
| 22 |
+
act: 'comp'
|
| 23 |
+
sample_rate: ${sample_rate}
|
| 24 |
+
num_bins: 1025
|
cfg/model/demucs.yaml
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
-
_target_: remfx.models.
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
@@ -13,4 +13,3 @@ model:
|
|
| 13 |
audio_channels: 1
|
| 14 |
nfft: 4096
|
| 15 |
sample_rate: ${sample_rate}
|
| 16 |
-
|
|
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
+
_target_: remfx.models.RemFX
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
|
|
| 13 |
audio_channels: 1
|
| 14 |
nfft: 4096
|
| 15 |
sample_rate: ${sample_rate}
|
|
|
cfg/model/dptnet.yaml
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
-
_target_: remfx.models.
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
@@ -16,3 +16,5 @@ model:
|
|
| 16 |
segment_size: 250
|
| 17 |
nspk: 1
|
| 18 |
win_len: 2
|
|
|
|
|
|
|
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
+
_target_: remfx.models.RemFX
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
|
|
| 16 |
segment_size: 250
|
| 17 |
nspk: 1
|
| 18 |
win_len: 2
|
| 19 |
+
sample_rate: ${sample_rate}
|
| 20 |
+
num_bins: 1025
|
cfg/model/umx.yaml
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
-
_target_: remfx.models.
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
@@ -14,4 +14,3 @@ model:
|
|
| 14 |
n_channels: 1
|
| 15 |
alpha: 0.3
|
| 16 |
sample_rate: ${sample_rate}
|
| 17 |
-
|
|
|
|
| 1 |
# @package _global_
|
| 2 |
model:
|
| 3 |
+
_target_: remfx.models.RemFX
|
| 4 |
lr: 1e-4
|
| 5 |
lr_beta1: 0.95
|
| 6 |
lr_beta2: 0.999
|
|
|
|
| 14 |
n_channels: 1
|
| 15 |
alpha: 0.3
|
| 16 |
sample_rate: ${sample_rate}
|
|
|
remfx/callbacks.py
ADDED
|
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from pytorch_lightning.callbacks import Callback
|
| 2 |
+
import pytorch_lightning as pl
|
| 3 |
+
from einops import rearrange
|
| 4 |
+
import torch
|
| 5 |
+
import wandb
|
| 6 |
+
from torch import Tensor
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class AudioCallback(Callback):
|
| 10 |
+
def __init__(self, sample_rate, *args, **kwargs):
|
| 11 |
+
super().__init__(*args, **kwargs)
|
| 12 |
+
self.log_train_audio = True
|
| 13 |
+
self.sample_rate = sample_rate
|
| 14 |
+
|
| 15 |
+
def on_train_batch_start(self, trainer, pl_module, batch, batch_idx):
|
| 16 |
+
# Log initial audio
|
| 17 |
+
if self.log_train_audio:
|
| 18 |
+
x, y, _, _ = batch
|
| 19 |
+
# Concat samples together for easier viewing in dashboard
|
| 20 |
+
input_samples = rearrange(x, "b c t -> c (b t)").unsqueeze(0)
|
| 21 |
+
target_samples = rearrange(y, "b c t -> c (b t)").unsqueeze(0)
|
| 22 |
+
|
| 23 |
+
log_wandb_audio_batch(
|
| 24 |
+
logger=trainer.logger,
|
| 25 |
+
id="input_effected_audio",
|
| 26 |
+
samples=input_samples.cpu(),
|
| 27 |
+
sampling_rate=self.sample_rate,
|
| 28 |
+
caption="Training Data",
|
| 29 |
+
)
|
| 30 |
+
log_wandb_audio_batch(
|
| 31 |
+
logger=trainer.logger,
|
| 32 |
+
id="target_audio",
|
| 33 |
+
samples=target_samples.cpu(),
|
| 34 |
+
sampling_rate=self.sample_rate,
|
| 35 |
+
caption="Target Data",
|
| 36 |
+
)
|
| 37 |
+
self.log_train_audio = False
|
| 38 |
+
|
| 39 |
+
def on_validation_batch_start(
|
| 40 |
+
self, trainer, pl_module, batch, batch_idx, dataloader_idx
|
| 41 |
+
):
|
| 42 |
+
x, target, _, _ = batch
|
| 43 |
+
# Only run on first batch
|
| 44 |
+
if batch_idx == 0:
|
| 45 |
+
with torch.no_grad():
|
| 46 |
+
y = pl_module.model.sample(x)
|
| 47 |
+
# Concat samples together for easier viewing in dashboard
|
| 48 |
+
# 2 seconds of silence between each sample
|
| 49 |
+
silence = torch.zeros_like(x)
|
| 50 |
+
silence = silence[:, : self.sample_rate * 2]
|
| 51 |
+
|
| 52 |
+
concat_samples = torch.cat([y, silence, x, silence, target], dim=-1)
|
| 53 |
+
log_wandb_audio_batch(
|
| 54 |
+
logger=trainer.logger,
|
| 55 |
+
id="prediction_input_target",
|
| 56 |
+
samples=concat_samples.cpu(),
|
| 57 |
+
sampling_rate=self.sample_rate,
|
| 58 |
+
caption=f"Epoch {trainer.current_epoch}",
|
| 59 |
+
)
|
| 60 |
+
|
| 61 |
+
def on_test_batch_start(self, *args):
|
| 62 |
+
self.on_validation_batch_start(*args)
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
class MetricCallback(Callback):
|
| 66 |
+
def on_validation_batch_start(
|
| 67 |
+
self, trainer, pl_module, batch, batch_idx, dataloader_idx
|
| 68 |
+
):
|
| 69 |
+
x, target, _, _ = batch
|
| 70 |
+
# Log Input Metrics
|
| 71 |
+
for metric in pl_module.metrics:
|
| 72 |
+
# SISDR returns negative values, so negate them
|
| 73 |
+
if metric == "SISDR":
|
| 74 |
+
negate = -1
|
| 75 |
+
else:
|
| 76 |
+
negate = 1
|
| 77 |
+
# Only Log FAD on test set
|
| 78 |
+
if metric == "FAD":
|
| 79 |
+
continue
|
| 80 |
+
pl_module.log(
|
| 81 |
+
f"Input_{metric}",
|
| 82 |
+
negate * pl_module.metrics[metric](x, target),
|
| 83 |
+
on_step=False,
|
| 84 |
+
on_epoch=True,
|
| 85 |
+
logger=True,
|
| 86 |
+
prog_bar=True,
|
| 87 |
+
sync_dist=True,
|
| 88 |
+
)
|
| 89 |
+
|
| 90 |
+
def on_test_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx):
|
| 91 |
+
self.on_validation_batch_start(
|
| 92 |
+
trainer, pl_module, batch, batch_idx, dataloader_idx
|
| 93 |
+
)
|
| 94 |
+
# Log FAD
|
| 95 |
+
x, target, _, _ = batch
|
| 96 |
+
pl_module.log(
|
| 97 |
+
"Input_FAD",
|
| 98 |
+
pl_module.metrics["FAD"](x, target),
|
| 99 |
+
on_step=False,
|
| 100 |
+
on_epoch=True,
|
| 101 |
+
logger=True,
|
| 102 |
+
prog_bar=True,
|
| 103 |
+
sync_dist=True,
|
| 104 |
+
)
|
| 105 |
+
|
| 106 |
+
|
| 107 |
+
def log_wandb_audio_batch(
|
| 108 |
+
logger: pl.loggers.WandbLogger,
|
| 109 |
+
id: str,
|
| 110 |
+
samples: Tensor,
|
| 111 |
+
sampling_rate: int,
|
| 112 |
+
caption: str = "",
|
| 113 |
+
max_items: int = 10,
|
| 114 |
+
):
|
| 115 |
+
num_items = samples.shape[0]
|
| 116 |
+
samples = rearrange(samples, "b c t -> b t c")
|
| 117 |
+
for idx in range(num_items):
|
| 118 |
+
if idx >= max_items:
|
| 119 |
+
break
|
| 120 |
+
logger.experiment.log(
|
| 121 |
+
{
|
| 122 |
+
f"{id}_{idx}": wandb.Audio(
|
| 123 |
+
samples[idx].cpu().numpy(),
|
| 124 |
+
caption=caption,
|
| 125 |
+
sample_rate=sampling_rate,
|
| 126 |
+
)
|
| 127 |
+
}
|
| 128 |
+
)
|
remfx/datasets.py
CHANGED
|
@@ -5,7 +5,6 @@ import torch
|
|
| 5 |
import shutil
|
| 6 |
import torchaudio
|
| 7 |
import pytorch_lightning as pl
|
| 8 |
-
import torch.nn.functional as F
|
| 9 |
|
| 10 |
from tqdm import tqdm
|
| 11 |
from pathlib import Path
|
|
|
|
| 5 |
import shutil
|
| 6 |
import torchaudio
|
| 7 |
import pytorch_lightning as pl
|
|
|
|
| 8 |
|
| 9 |
from tqdm import tqdm
|
| 10 |
from pathlib import Path
|
remfx/dcunet.py
CHANGED
|
@@ -5,11 +5,11 @@ import torch
|
|
| 5 |
import torch.nn as nn
|
| 6 |
import torch.nn.functional as F
|
| 7 |
import numpy as np
|
| 8 |
-
from utils import single, concat_complex
|
| 9 |
from torch.nn.init import calculate_gain
|
| 10 |
from typing import Tuple
|
| 11 |
from scipy.signal import get_window
|
| 12 |
from librosa.util import pad_center
|
|
|
|
| 13 |
|
| 14 |
|
| 15 |
class ComplexConvBlock(nn.Module):
|
|
@@ -549,7 +549,7 @@ class ComplexActLayer(nn.Module):
|
|
| 549 |
|
| 550 |
def forward(self, x):
|
| 551 |
real, img = x.chunk(2, 1)
|
| 552 |
-
return torch.cat([F.
|
| 553 |
|
| 554 |
|
| 555 |
class STFT(nn.Module):
|
|
|
|
| 5 |
import torch.nn as nn
|
| 6 |
import torch.nn.functional as F
|
| 7 |
import numpy as np
|
|
|
|
| 8 |
from torch.nn.init import calculate_gain
|
| 9 |
from typing import Tuple
|
| 10 |
from scipy.signal import get_window
|
| 11 |
from librosa.util import pad_center
|
| 12 |
+
from remfx.utils import single, concat_complex
|
| 13 |
|
| 14 |
|
| 15 |
class ComplexConvBlock(nn.Module):
|
|
|
|
| 549 |
|
| 550 |
def forward(self, x):
|
| 551 |
real, img = x.chunk(2, 1)
|
| 552 |
+
return torch.cat([F.leaky_relu(real), torch.tanh(img) * np.pi], dim=1)
|
| 553 |
|
| 554 |
|
| 555 |
class STFT(nn.Module):
|
remfx/dptnet.py
CHANGED
|
@@ -57,11 +57,10 @@ class DPTNet_base(nn.Module):
|
|
| 57 |
self.mask_conv1x1 = nn.Conv1d(self.feature_dim, self.enc_dim, 1, bias=False)
|
| 58 |
self.decoder = DPTDecoder(n_filters=enc_dim, window_size=win_len)
|
| 59 |
|
| 60 |
-
def forward(self,
|
| 61 |
"""
|
| 62 |
mix: shape (batch, T)
|
| 63 |
"""
|
| 64 |
-
mix, target = batch
|
| 65 |
batch_size = mix.shape[0]
|
| 66 |
mix = self.dpt_encoder(mix) # (B, E, L)
|
| 67 |
|
|
|
|
| 57 |
self.mask_conv1x1 = nn.Conv1d(self.feature_dim, self.enc_dim, 1, bias=False)
|
| 58 |
self.decoder = DPTDecoder(n_filters=enc_dim, window_size=win_len)
|
| 59 |
|
| 60 |
+
def forward(self, mix):
|
| 61 |
"""
|
| 62 |
mix: shape (batch, T)
|
| 63 |
"""
|
|
|
|
| 64 |
batch_size = mix.shape[0]
|
| 65 |
mix = self.dpt_encoder(mix) # (B, E, L)
|
| 66 |
|
remfx/models.py
CHANGED
|
@@ -2,16 +2,16 @@ import torch
|
|
| 2 |
import torchmetrics
|
| 3 |
import pytorch_lightning as pl
|
| 4 |
from torch import Tensor, nn
|
| 5 |
-
from
|
| 6 |
from torchaudio.models import HDemucs
|
| 7 |
from audio_diffusion_pytorch import DiffusionModel
|
| 8 |
from auraloss.time import SISDRLoss
|
| 9 |
from auraloss.freq import MultiResolutionSTFTLoss
|
| 10 |
from umx.openunmix.model import OpenUnmix, Separator
|
| 11 |
|
| 12 |
-
from utils import FADLoss, spectrogram
|
| 13 |
-
from dptnet import DPTNet_base
|
| 14 |
-
from dcunet import RefineSpectrogramUnet
|
| 15 |
|
| 16 |
|
| 17 |
class RemFX(pl.LightningModule):
|
|
@@ -55,41 +55,29 @@ class RemFX(pl.LightningModule):
|
|
| 55 |
eps=self.lr_eps,
|
| 56 |
weight_decay=self.lr_weight_decay,
|
| 57 |
)
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
optimizer.step(closure=optimizer_closure)
|
| 73 |
-
|
| 74 |
-
# update learning rate. Reduce by factor of 10 at 80% and 95% of training
|
| 75 |
-
if self.trainer.global_step == 0.8 * self.trainer.max_steps:
|
| 76 |
-
for pg in optimizer.param_groups:
|
| 77 |
-
pg["lr"] = 0.1 * pg["lr"]
|
| 78 |
-
if self.trainer.global_step == 0.95 * self.trainer.max_steps:
|
| 79 |
-
for pg in optimizer.param_groups:
|
| 80 |
-
pg["lr"] = 0.1 * pg["lr"]
|
| 81 |
|
| 82 |
def training_step(self, batch, batch_idx):
|
| 83 |
-
|
| 84 |
-
return loss
|
| 85 |
|
| 86 |
def validation_step(self, batch, batch_idx):
|
| 87 |
-
|
| 88 |
-
return loss
|
| 89 |
|
| 90 |
def test_step(self, batch, batch_idx):
|
| 91 |
-
|
| 92 |
-
return loss
|
| 93 |
|
| 94 |
def common_step(self, batch, batch_idx, mode: str = "train"):
|
| 95 |
x, y, _, _ = batch # x, y = (B, C, T), (B, C, T)
|
|
@@ -116,89 +104,8 @@ class RemFX(pl.LightningModule):
|
|
| 116 |
prog_bar=True,
|
| 117 |
sync_dist=True,
|
| 118 |
)
|
| 119 |
-
|
| 120 |
return loss
|
| 121 |
|
| 122 |
-
def on_train_batch_start(self, batch, batch_idx):
|
| 123 |
-
# Log initial audio
|
| 124 |
-
if self.log_train_audio:
|
| 125 |
-
x, y, _, _ = batch
|
| 126 |
-
# Concat samples together for easier viewing in dashboard
|
| 127 |
-
input_samples = rearrange(x, "b c t -> c (b t)").unsqueeze(0)
|
| 128 |
-
target_samples = rearrange(y, "b c t -> c (b t)").unsqueeze(0)
|
| 129 |
-
|
| 130 |
-
log_wandb_audio_batch(
|
| 131 |
-
logger=self.logger,
|
| 132 |
-
id="input_effected_audio",
|
| 133 |
-
samples=input_samples.cpu(),
|
| 134 |
-
sampling_rate=self.sample_rate,
|
| 135 |
-
caption="Training Data",
|
| 136 |
-
)
|
| 137 |
-
log_wandb_audio_batch(
|
| 138 |
-
logger=self.logger,
|
| 139 |
-
id="target_audio",
|
| 140 |
-
samples=target_samples.cpu(),
|
| 141 |
-
sampling_rate=self.sample_rate,
|
| 142 |
-
caption="Target Data",
|
| 143 |
-
)
|
| 144 |
-
self.log_train_audio = False
|
| 145 |
-
|
| 146 |
-
def on_validation_batch_start(self, batch, batch_idx, dataloader_idx):
|
| 147 |
-
x, target, _, _ = batch
|
| 148 |
-
# Log Input Metrics
|
| 149 |
-
for metric in self.metrics:
|
| 150 |
-
# SISDR returns negative values, so negate them
|
| 151 |
-
if metric == "SISDR":
|
| 152 |
-
negate = -1
|
| 153 |
-
else:
|
| 154 |
-
negate = 1
|
| 155 |
-
# Only Log FAD on test set
|
| 156 |
-
if metric == "FAD":
|
| 157 |
-
continue
|
| 158 |
-
self.log(
|
| 159 |
-
f"Input_{metric}",
|
| 160 |
-
negate * self.metrics[metric](x, target),
|
| 161 |
-
on_step=False,
|
| 162 |
-
on_epoch=True,
|
| 163 |
-
logger=True,
|
| 164 |
-
prog_bar=True,
|
| 165 |
-
sync_dist=True,
|
| 166 |
-
)
|
| 167 |
-
# Only run on first batch
|
| 168 |
-
if batch_idx == 0:
|
| 169 |
-
self.model.eval()
|
| 170 |
-
with torch.no_grad():
|
| 171 |
-
y = self.model.sample(x)
|
| 172 |
-
|
| 173 |
-
# Concat samples together for easier viewing in dashboard
|
| 174 |
-
# 2 seconds of silence between each sample
|
| 175 |
-
silence = torch.zeros_like(x)
|
| 176 |
-
silence = silence[:, : self.sample_rate * 2]
|
| 177 |
-
|
| 178 |
-
concat_samples = torch.cat([y, silence, x, silence, target], dim=-1)
|
| 179 |
-
log_wandb_audio_batch(
|
| 180 |
-
logger=self.logger,
|
| 181 |
-
id="prediction_input_target",
|
| 182 |
-
samples=concat_samples.cpu(),
|
| 183 |
-
sampling_rate=self.sample_rate,
|
| 184 |
-
caption=f"Epoch {self.current_epoch}",
|
| 185 |
-
)
|
| 186 |
-
self.model.train()
|
| 187 |
-
|
| 188 |
-
def on_test_batch_start(self, batch, batch_idx, dataloader_idx):
|
| 189 |
-
self.on_validation_batch_start(batch, batch_idx, dataloader_idx)
|
| 190 |
-
# Log FAD
|
| 191 |
-
x, target, _, _ = batch
|
| 192 |
-
self.log(
|
| 193 |
-
"Input_FAD",
|
| 194 |
-
self.metrics["FAD"](x, target),
|
| 195 |
-
on_step=False,
|
| 196 |
-
on_epoch=True,
|
| 197 |
-
logger=True,
|
| 198 |
-
prog_bar=True,
|
| 199 |
-
sync_dist=True,
|
| 200 |
-
)
|
| 201 |
-
|
| 202 |
|
| 203 |
class OpenUnmixModel(nn.Module):
|
| 204 |
def __init__(
|
|
@@ -284,9 +191,10 @@ class DiffusionGenerationModel(nn.Module):
|
|
| 284 |
|
| 285 |
|
| 286 |
class DPTNetModel(nn.Module):
|
| 287 |
-
def __init__(self, sample_rate, **kwargs):
|
| 288 |
super().__init__()
|
| 289 |
self.model = DPTNet_base(**kwargs)
|
|
|
|
| 290 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
| 291 |
n_bins=self.num_bins, sample_rate=sample_rate
|
| 292 |
)
|
|
@@ -294,31 +202,42 @@ class DPTNetModel(nn.Module):
|
|
| 294 |
|
| 295 |
def forward(self, batch):
|
| 296 |
x, target = batch
|
| 297 |
-
output = self.model(x
|
| 298 |
loss = self.mrstftloss(output, target) + self.l1loss(output, target) * 100
|
| 299 |
return loss, output
|
| 300 |
|
| 301 |
def sample(self, x: Tensor) -> Tensor:
|
| 302 |
-
return self.model.
|
| 303 |
|
| 304 |
|
| 305 |
class DCUNetModel(nn.Module):
|
| 306 |
-
def __init__(self, sample_rate, **kwargs):
|
| 307 |
super().__init__()
|
| 308 |
self.model = RefineSpectrogramUnet(**kwargs)
|
| 309 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
| 310 |
-
n_bins=
|
| 311 |
)
|
| 312 |
self.l1loss = nn.L1Loss()
|
| 313 |
|
| 314 |
def forward(self, batch):
|
| 315 |
x, target = batch
|
| 316 |
-
output = self.model(x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 317 |
loss = self.mrstftloss(output, target) + self.l1loss(output, target) * 100
|
| 318 |
return loss, output
|
| 319 |
|
| 320 |
def sample(self, x: Tensor) -> Tensor:
|
| 321 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 322 |
|
| 323 |
|
| 324 |
class FXClassifier(pl.LightningModule):
|
|
|
|
| 2 |
import torchmetrics
|
| 3 |
import pytorch_lightning as pl
|
| 4 |
from torch import Tensor, nn
|
| 5 |
+
from torch.nn import functional as F
|
| 6 |
from torchaudio.models import HDemucs
|
| 7 |
from audio_diffusion_pytorch import DiffusionModel
|
| 8 |
from auraloss.time import SISDRLoss
|
| 9 |
from auraloss.freq import MultiResolutionSTFTLoss
|
| 10 |
from umx.openunmix.model import OpenUnmix, Separator
|
| 11 |
|
| 12 |
+
from remfx.utils import FADLoss, spectrogram
|
| 13 |
+
from remfx.dptnet import DPTNet_base
|
| 14 |
+
from remfx.dcunet import RefineSpectrogramUnet
|
| 15 |
|
| 16 |
|
| 17 |
class RemFX(pl.LightningModule):
|
|
|
|
| 55 |
eps=self.lr_eps,
|
| 56 |
weight_decay=self.lr_weight_decay,
|
| 57 |
)
|
| 58 |
+
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
|
| 59 |
+
optimizer,
|
| 60 |
+
[0.8 * self.trainer.max_steps, 0.95 * self.trainer.max_steps],
|
| 61 |
+
gamma=0.1,
|
| 62 |
+
)
|
| 63 |
+
return {
|
| 64 |
+
"optimizer": optimizer,
|
| 65 |
+
"lr_scheduler": {
|
| 66 |
+
"scheduler": lr_scheduler,
|
| 67 |
+
"monitor": "val_loss",
|
| 68 |
+
"interval": "step",
|
| 69 |
+
"frequency": 1,
|
| 70 |
+
},
|
| 71 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 72 |
|
| 73 |
def training_step(self, batch, batch_idx):
|
| 74 |
+
return self.common_step(batch, batch_idx, mode="train")
|
|
|
|
| 75 |
|
| 76 |
def validation_step(self, batch, batch_idx):
|
| 77 |
+
return self.common_step(batch, batch_idx, mode="valid")
|
|
|
|
| 78 |
|
| 79 |
def test_step(self, batch, batch_idx):
|
| 80 |
+
return self.common_step(batch, batch_idx, mode="test")
|
|
|
|
| 81 |
|
| 82 |
def common_step(self, batch, batch_idx, mode: str = "train"):
|
| 83 |
x, y, _, _ = batch # x, y = (B, C, T), (B, C, T)
|
|
|
|
| 104 |
prog_bar=True,
|
| 105 |
sync_dist=True,
|
| 106 |
)
|
|
|
|
| 107 |
return loss
|
| 108 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 109 |
|
| 110 |
class OpenUnmixModel(nn.Module):
|
| 111 |
def __init__(
|
|
|
|
| 191 |
|
| 192 |
|
| 193 |
class DPTNetModel(nn.Module):
|
| 194 |
+
def __init__(self, sample_rate, num_bins, **kwargs):
|
| 195 |
super().__init__()
|
| 196 |
self.model = DPTNet_base(**kwargs)
|
| 197 |
+
self.num_bins = num_bins
|
| 198 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
| 199 |
n_bins=self.num_bins, sample_rate=sample_rate
|
| 200 |
)
|
|
|
|
| 202 |
|
| 203 |
def forward(self, batch):
|
| 204 |
x, target = batch
|
| 205 |
+
output = self.model(x.squeeze(1))
|
| 206 |
loss = self.mrstftloss(output, target) + self.l1loss(output, target) * 100
|
| 207 |
return loss, output
|
| 208 |
|
| 209 |
def sample(self, x: Tensor) -> Tensor:
|
| 210 |
+
return self.model(x.squeeze(1))
|
| 211 |
|
| 212 |
|
| 213 |
class DCUNetModel(nn.Module):
|
| 214 |
+
def __init__(self, sample_rate, num_bins, **kwargs):
|
| 215 |
super().__init__()
|
| 216 |
self.model = RefineSpectrogramUnet(**kwargs)
|
| 217 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
| 218 |
+
n_bins=num_bins, sample_rate=sample_rate
|
| 219 |
)
|
| 220 |
self.l1loss = nn.L1Loss()
|
| 221 |
|
| 222 |
def forward(self, batch):
|
| 223 |
x, target = batch
|
| 224 |
+
output = self.model(x.squeeze(1)).unsqueeze(1) # B x 1 x T
|
| 225 |
+
# Pad or crop to match target
|
| 226 |
+
if output.shape[-1] > target.shape[-1]:
|
| 227 |
+
output = output[:, : target.shape[-1]]
|
| 228 |
+
elif output.shape[-1] < target.shape[-1]:
|
| 229 |
+
output = F.pad(output, (0, target.shape[-1] - output.shape[-1]))
|
| 230 |
loss = self.mrstftloss(output, target) + self.l1loss(output, target) * 100
|
| 231 |
return loss, output
|
| 232 |
|
| 233 |
def sample(self, x: Tensor) -> Tensor:
|
| 234 |
+
output = self.model(x.squeeze(1)).unsqueeze(1) # B x 1 x T
|
| 235 |
+
# Pad or crop to match target
|
| 236 |
+
if output.shape[-1] > x.shape[-1]:
|
| 237 |
+
output = output[:, : x.shape[-1]]
|
| 238 |
+
elif output.shape[-1] < x.shape[-1]:
|
| 239 |
+
output = F.pad(output, (0, x.shape[-1] - output.shape[-1]))
|
| 240 |
+
return output
|
| 241 |
|
| 242 |
|
| 243 |
class FXClassifier(pl.LightningModule):
|
remfx/utils.py
CHANGED
|
@@ -7,10 +7,8 @@ from frechet_audio_distance import FrechetAudioDistance
|
|
| 7 |
import numpy as np
|
| 8 |
import torch
|
| 9 |
import torchaudio
|
| 10 |
-
from torch import
|
| 11 |
-
import
|
| 12 |
-
from einops import rearrange
|
| 13 |
-
from torch._six import container_abcs
|
| 14 |
|
| 15 |
|
| 16 |
def get_logger(name=__name__) -> logging.Logger:
|
|
@@ -144,30 +142,6 @@ def create_sequential_chunks(
|
|
| 144 |
return chunks, sr
|
| 145 |
|
| 146 |
|
| 147 |
-
def log_wandb_audio_batch(
|
| 148 |
-
logger: pl.loggers.WandbLogger,
|
| 149 |
-
id: str,
|
| 150 |
-
samples: Tensor,
|
| 151 |
-
sampling_rate: int,
|
| 152 |
-
caption: str = "",
|
| 153 |
-
max_items: int = 10,
|
| 154 |
-
):
|
| 155 |
-
num_items = samples.shape[0]
|
| 156 |
-
samples = rearrange(samples, "b c t -> b t c")
|
| 157 |
-
for idx in range(num_items):
|
| 158 |
-
if idx >= max_items:
|
| 159 |
-
break
|
| 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,
|
|
@@ -209,7 +183,7 @@ def init_bn(bn):
|
|
| 209 |
|
| 210 |
def _ntuple(n: int):
|
| 211 |
def parse(x):
|
| 212 |
-
if isinstance(x,
|
| 213 |
return x
|
| 214 |
return tuple([x] * n)
|
| 215 |
|
|
|
|
| 7 |
import numpy as np
|
| 8 |
import torch
|
| 9 |
import torchaudio
|
| 10 |
+
from torch import nn
|
| 11 |
+
import collections.abc
|
|
|
|
|
|
|
| 12 |
|
| 13 |
|
| 14 |
def get_logger(name=__name__) -> logging.Logger:
|
|
|
|
| 142 |
return chunks, sr
|
| 143 |
|
| 144 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
def spectrogram(
|
| 146 |
x: torch.Tensor,
|
| 147 |
window: torch.Tensor,
|
|
|
|
| 183 |
|
| 184 |
def _ntuple(n: int):
|
| 185 |
def parse(x):
|
| 186 |
+
if isinstance(x, collections.abc.Iterable):
|
| 187 |
return x
|
| 188 |
return tuple([x] * n)
|
| 189 |
|