Spaces:
Sleeping
Sleeping
Commit
·
943f213
1
Parent(s):
d92486b
Fix loss issue in chain_inference
Browse files- remfx/models.py +58 -22
remfx/models.py
CHANGED
@@ -12,6 +12,7 @@ from remfx.utils import FADLoss, spectrogram
|
|
12 |
from remfx.tcn import TCN
|
13 |
from remfx.utils import causal_crop
|
14 |
from remfx.callbacks import log_wandb_audio_batch
|
|
|
15 |
from remfx import effects
|
16 |
import asteroid
|
17 |
|
@@ -47,6 +48,23 @@ class RemFXChainInference(pl.LightningModule):
|
|
47 |
for effect_label in rem_fx_labels
|
48 |
]
|
49 |
output = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
with torch.no_grad():
|
51 |
for i, (elem, effects_list) in enumerate(zip(x, effects)):
|
52 |
elem = elem.unsqueeze(0) # Add batch dim
|
@@ -56,33 +74,41 @@ class RemFXChainInference(pl.LightningModule):
|
|
56 |
effect for effect in effects_order if effect in effect_list_names
|
57 |
]
|
58 |
|
59 |
-
log_wandb_audio_batch(
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
)
|
66 |
for effect in effects:
|
67 |
# Sample the model
|
68 |
elem = self.model[effect].model.sample(elem)
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
log_wandb_audio_batch(
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
)
|
83 |
output.append(elem.squeeze(0))
|
84 |
output = torch.stack(output)
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
loss = self.mrstftloss(output, y) + self.l1loss(output, y) * 100
|
87 |
return loss, output
|
88 |
|
@@ -112,6 +138,16 @@ class RemFXChainInference(pl.LightningModule):
|
|
112 |
prog_bar=True,
|
113 |
sync_dist=True,
|
114 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
|
116 |
def sample(self, batch):
|
117 |
return self.forward(batch, 0)[1]
|
|
|
12 |
from remfx.tcn import TCN
|
13 |
from remfx.utils import causal_crop
|
14 |
from remfx.callbacks import log_wandb_audio_batch
|
15 |
+
from einops import rearrange
|
16 |
from remfx import effects
|
17 |
import asteroid
|
18 |
|
|
|
48 |
for effect_label in rem_fx_labels
|
49 |
]
|
50 |
output = []
|
51 |
+
input_samples = rearrange(x, "b c t -> c (b t)").unsqueeze(0)
|
52 |
+
target_samples = rearrange(y, "b c t -> c (b t)").unsqueeze(0)
|
53 |
+
|
54 |
+
log_wandb_audio_batch(
|
55 |
+
logger=self.logger,
|
56 |
+
id="input_effected_audio",
|
57 |
+
samples=input_samples.cpu(),
|
58 |
+
sampling_rate=self.sample_rate,
|
59 |
+
caption=effects,
|
60 |
+
)
|
61 |
+
log_wandb_audio_batch(
|
62 |
+
logger=self.logger,
|
63 |
+
id="target_audio",
|
64 |
+
samples=target_samples.cpu(),
|
65 |
+
sampling_rate=self.sample_rate,
|
66 |
+
caption="Target Data",
|
67 |
+
)
|
68 |
with torch.no_grad():
|
69 |
for i, (elem, effects_list) in enumerate(zip(x, effects)):
|
70 |
elem = elem.unsqueeze(0) # Add batch dim
|
|
|
74 |
effect for effect in effects_order if effect in effect_list_names
|
75 |
]
|
76 |
|
77 |
+
# log_wandb_audio_batch(
|
78 |
+
# logger=self.logger,
|
79 |
+
# id=f"{i}_Before",
|
80 |
+
# samples=elem.cpu(),
|
81 |
+
# sampling_rate=self.sample_rate,
|
82 |
+
# caption=effects,
|
83 |
+
# )
|
84 |
for effect in effects:
|
85 |
# Sample the model
|
86 |
elem = self.model[effect].model.sample(elem)
|
87 |
+
# log_wandb_audio_batch(
|
88 |
+
# logger=self.logger,
|
89 |
+
# id=f"{i}_{effect}",
|
90 |
+
# samples=elem.cpu(),
|
91 |
+
# sampling_rate=self.sample_rate,
|
92 |
+
# caption=effects,
|
93 |
+
# )
|
94 |
+
# log_wandb_audio_batch(
|
95 |
+
# logger=self.logger,
|
96 |
+
# id=f"{i}_After",
|
97 |
+
# samples=elem.cpu(),
|
98 |
+
# sampling_rate=self.sample_rate,
|
99 |
+
# caption=effects,
|
100 |
+
# )
|
101 |
output.append(elem.squeeze(0))
|
102 |
output = torch.stack(output)
|
103 |
+
output_samples = rearrange(output, "b c t -> c (b t)").unsqueeze(0)
|
104 |
+
|
105 |
+
log_wandb_audio_batch(
|
106 |
+
logger=self.logger,
|
107 |
+
id="output_audio",
|
108 |
+
samples=output_samples.cpu(),
|
109 |
+
sampling_rate=self.sample_rate,
|
110 |
+
caption="Output Data",
|
111 |
+
)
|
112 |
loss = self.mrstftloss(output, y) + self.l1loss(output, y) * 100
|
113 |
return loss, output
|
114 |
|
|
|
138 |
prog_bar=True,
|
139 |
sync_dist=True,
|
140 |
)
|
141 |
+
self.log(
|
142 |
+
f"Input_{metric}",
|
143 |
+
negate * self.metrics[metric](x, y),
|
144 |
+
on_step=False,
|
145 |
+
on_epoch=True,
|
146 |
+
logger=True,
|
147 |
+
prog_bar=True,
|
148 |
+
sync_dist=True,
|
149 |
+
)
|
150 |
+
return loss
|
151 |
|
152 |
def sample(self, batch):
|
153 |
return self.forward(batch, 0)[1]
|