Spaces:
Sleeping
Sleeping
Commit
·
ff526b3
1
Parent(s):
1530829
Add metrics
Browse files- remfx/models.py +33 -21
remfx/models.py
CHANGED
|
@@ -4,7 +4,9 @@ import pytorch_lightning as pl
|
|
| 4 |
from einops import rearrange
|
| 5 |
import wandb
|
| 6 |
from audio_diffusion_pytorch import DiffusionModel
|
| 7 |
-
import
|
|
|
|
|
|
|
| 8 |
|
| 9 |
from umx.openunmix.model import OpenUnmix, Separator
|
| 10 |
|
|
@@ -28,6 +30,13 @@ class RemFXModel(pl.LightningModule):
|
|
| 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):
|
|
@@ -49,10 +58,23 @@ class RemFXModel(pl.LightningModule):
|
|
| 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):
|
|
@@ -62,24 +84,13 @@ class RemFXModel(pl.LightningModule):
|
|
| 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="
|
| 68 |
-
samples=
|
| 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 |
)
|
|
@@ -116,7 +127,7 @@ class OpenUnmixModel(torch.nn.Module):
|
|
| 116 |
n_fft=self.n_fft,
|
| 117 |
n_hop=self.hop_length,
|
| 118 |
)
|
| 119 |
-
self.loss_fn =
|
| 120 |
n_bins=self.num_bins, sample_rate=self.sample_rate
|
| 121 |
)
|
| 122 |
|
|
@@ -127,7 +138,7 @@ class OpenUnmixModel(torch.nn.Module):
|
|
| 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)
|
|
@@ -140,7 +151,8 @@ class DiffusionGenerationModel(nn.Module):
|
|
| 140 |
|
| 141 |
def forward(self, batch):
|
| 142 |
x, target, label = batch
|
| 143 |
-
|
|
|
|
| 144 |
|
| 145 |
def sample(self, x: Tensor, num_steps: int = 10) -> Tensor:
|
| 146 |
noise = torch.randn(x.shape).to(x)
|
|
|
|
| 4 |
from einops import rearrange
|
| 5 |
import wandb
|
| 6 |
from audio_diffusion_pytorch import DiffusionModel
|
| 7 |
+
from auraloss.time import SISDRLoss
|
| 8 |
+
from auraloss.freq import MultiResolutionSTFTLoss, STFTLoss
|
| 9 |
+
from torch.nn import L1Loss
|
| 10 |
|
| 11 |
from umx.openunmix.model import OpenUnmix, Separator
|
| 12 |
|
|
|
|
| 30 |
self.lr_weight_decay = lr_weight_decay
|
| 31 |
self.sample_rate = sample_rate
|
| 32 |
self.model = network
|
| 33 |
+
self.metrics = torch.nn.ModuleDict(
|
| 34 |
+
{
|
| 35 |
+
"SISDR": SISDRLoss(),
|
| 36 |
+
"STFT": STFTLoss(),
|
| 37 |
+
"L1": L1Loss(),
|
| 38 |
+
}
|
| 39 |
+
)
|
| 40 |
|
| 41 |
@property
|
| 42 |
def device(self):
|
|
|
|
| 58 |
|
| 59 |
def validation_step(self, batch, batch_idx):
|
| 60 |
loss = self.common_step(batch, batch_idx, mode="valid")
|
| 61 |
+
return loss
|
| 62 |
|
| 63 |
def common_step(self, batch, batch_idx, mode: str = "train"):
|
| 64 |
+
loss, output = self.model(batch)
|
| 65 |
self.log(f"{mode}_loss", loss)
|
| 66 |
+
x, y, label = batch
|
| 67 |
+
# Metric logging
|
| 68 |
+
for metric in self.metrics:
|
| 69 |
+
self.log(
|
| 70 |
+
f"{mode}_{metric}",
|
| 71 |
+
self.metrics[metric](output, y),
|
| 72 |
+
on_step=False,
|
| 73 |
+
on_epoch=True,
|
| 74 |
+
logger=True,
|
| 75 |
+
prog_bar=True,
|
| 76 |
+
)
|
| 77 |
+
|
| 78 |
return loss
|
| 79 |
|
| 80 |
def on_validation_epoch_start(self):
|
|
|
|
| 84 |
if self.log_next:
|
| 85 |
x, target, label = batch
|
| 86 |
y = self.model.sample(x)
|
| 87 |
+
|
| 88 |
+
# Concat samples together for easier viewing in dashboard
|
| 89 |
+
concat_samples = torch.cat([x, y, target], dim=-1)
|
| 90 |
log_wandb_audio_batch(
|
| 91 |
logger=self.logger,
|
| 92 |
+
id="prediction_sample_target",
|
| 93 |
+
samples=concat_samples.cpu(),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
sampling_rate=self.sample_rate,
|
| 95 |
caption=f"Epoch {self.current_epoch}",
|
| 96 |
)
|
|
|
|
| 127 |
n_fft=self.n_fft,
|
| 128 |
n_hop=self.hop_length,
|
| 129 |
)
|
| 130 |
+
self.loss_fn = MultiResolutionSTFTLoss(
|
| 131 |
n_bins=self.num_bins, sample_rate=self.sample_rate
|
| 132 |
)
|
| 133 |
|
|
|
|
| 138 |
sep_out = self.separator(x).squeeze(1)
|
| 139 |
loss = self.loss_fn(sep_out, target)
|
| 140 |
|
| 141 |
+
return loss, sep_out
|
| 142 |
|
| 143 |
def sample(self, x: Tensor) -> Tensor:
|
| 144 |
return self.separator(x).squeeze(1)
|
|
|
|
| 151 |
|
| 152 |
def forward(self, batch):
|
| 153 |
x, target, label = batch
|
| 154 |
+
sampled_out = self.model.sample(x)
|
| 155 |
+
return self.model(x), sampled_out
|
| 156 |
|
| 157 |
def sample(self, x: Tensor, num_steps: int = 10) -> Tensor:
|
| 158 |
noise = torch.randn(x.shape).to(x)
|