Spaces:
Sleeping
Sleeping
Commit
·
9eba2f5
1
Parent(s):
80a1624
Fix circular import issue
Browse files- cfg/exp/chain_inference.yaml +5 -3
- remfx/callbacks.py +3 -1
- remfx/models.py +34 -23
- scripts/chain_inference.py +5 -2
cfg/exp/chain_inference.yaml
CHANGED
@@ -6,7 +6,7 @@ seed: 12345
|
|
6 |
sample_rate: 48000
|
7 |
chunk_size: 262144 # 5.5s
|
8 |
logs_dir: "./logs"
|
9 |
-
render_files:
|
10 |
render_root: "/scratch/EffectSet"
|
11 |
accelerator: "gpu"
|
12 |
log_audio: True
|
@@ -26,10 +26,12 @@ effects_to_remove:
|
|
26 |
datamodule:
|
27 |
batch_size: 16
|
28 |
num_workers: 8
|
|
|
|
|
29 |
ckpts:
|
30 |
-
RandomPedalboardChorus: "ckpts/chorus.ckpt"
|
31 |
-
RandomPedalboardDelay: "ckpts/delay.ckpt"
|
32 |
RandomPedalboardDistortion: "ckpts/distortion.ckpt"
|
33 |
RandomPedalboardCompressor: "ckpts/compressor.ckpt"
|
34 |
RandomPedalboardReverb: "ckpts/reverb.ckpt"
|
|
|
|
|
35 |
num_bins: 1025
|
|
|
6 |
sample_rate: 48000
|
7 |
chunk_size: 262144 # 5.5s
|
8 |
logs_dir: "./logs"
|
9 |
+
render_files: False
|
10 |
render_root: "/scratch/EffectSet"
|
11 |
accelerator: "gpu"
|
12 |
log_audio: True
|
|
|
26 |
datamodule:
|
27 |
batch_size: 16
|
28 |
num_workers: 8
|
29 |
+
train_dataset: None
|
30 |
+
val_dataset: None
|
31 |
ckpts:
|
|
|
|
|
32 |
RandomPedalboardDistortion: "ckpts/distortion.ckpt"
|
33 |
RandomPedalboardCompressor: "ckpts/compressor.ckpt"
|
34 |
RandomPedalboardReverb: "ckpts/reverb.ckpt"
|
35 |
+
RandomPedalboardChorus: "ckpts/chorus.ckpt"
|
36 |
+
RandomPedalboardDelay: "ckpts/delay.ckpt"
|
37 |
num_bins: 1025
|
remfx/callbacks.py
CHANGED
@@ -4,7 +4,6 @@ from einops import rearrange
|
|
4 |
import torch
|
5 |
import wandb
|
6 |
from torch import Tensor
|
7 |
-
from remfx.models import RemFXChainInference
|
8 |
|
9 |
|
10 |
class AudioCallback(Callback):
|
@@ -47,6 +46,9 @@ class AudioCallback(Callback):
|
|
47 |
# Only run on first batch
|
48 |
if batch_idx == 0 and self.log_audio:
|
49 |
with torch.no_grad():
|
|
|
|
|
|
|
50 |
if type(pl_module) == RemFXChainInference:
|
51 |
y = pl_module.sample(batch)
|
52 |
else:
|
|
|
4 |
import torch
|
5 |
import wandb
|
6 |
from torch import Tensor
|
|
|
7 |
|
8 |
|
9 |
class AudioCallback(Callback):
|
|
|
46 |
# Only run on first batch
|
47 |
if batch_idx == 0 and self.log_audio:
|
48 |
with torch.no_grad():
|
49 |
+
# Avoids circular import
|
50 |
+
from remfx.models import RemFXChainInference
|
51 |
+
|
52 |
if type(pl_module) == RemFXChainInference:
|
53 |
y = pl_module.sample(batch)
|
54 |
else:
|
remfx/models.py
CHANGED
@@ -14,12 +14,13 @@ from remfx.utils import causal_crop
|
|
14 |
from remfx.callbacks import log_wandb_audio_batch
|
15 |
from remfx import effects
|
16 |
import asteroid
|
|
|
17 |
|
18 |
ALL_EFFECTS = effects.Pedalboard_Effects
|
19 |
|
20 |
|
21 |
class RemFXChainInference(pl.LightningModule):
|
22 |
-
def __init__(self, models, sample_rate, num_bins):
|
23 |
super().__init__()
|
24 |
self.model = models
|
25 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
@@ -33,36 +34,45 @@ class RemFXChainInference(pl.LightningModule):
|
|
33 |
"FAD": FADLoss(sample_rate=sample_rate),
|
34 |
}
|
35 |
)
|
|
|
|
|
36 |
|
37 |
-
def forward(self, batch):
|
38 |
x, y, _, rem_fx_labels = batch
|
39 |
# Use chain of effects defined in config
|
|
|
|
|
|
|
|
|
40 |
effects = [
|
41 |
[ALL_EFFECTS[i] for i, effect in enumerate(effect_label) if effect == 1.0]
|
42 |
for effect_label in rem_fx_labels
|
43 |
]
|
44 |
output = []
|
45 |
with torch.no_grad():
|
46 |
-
for i, (elem,
|
47 |
elem = elem.unsqueeze(0) # Add batch dim
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
)
|
|
|
|
|
|
|
|
|
55 |
for effect in effect_chain:
|
56 |
-
#
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
)
|
66 |
output.append(elem.squeeze(0))
|
67 |
output = torch.stack(output)
|
68 |
|
@@ -71,8 +81,9 @@ class RemFXChainInference(pl.LightningModule):
|
|
71 |
|
72 |
def test_step(self, batch, batch_idx):
|
73 |
x, y, _, _ = batch # x, y = (B, C, T), (B, C, T)
|
74 |
-
|
75 |
-
|
|
|
76 |
# Crop target to match output
|
77 |
if output.shape[-1] < y.shape[-1]:
|
78 |
y = causal_crop(y, output.shape[-1])
|
@@ -86,7 +97,7 @@ class RemFXChainInference(pl.LightningModule):
|
|
86 |
else:
|
87 |
negate = 1
|
88 |
self.log(
|
89 |
-
f"test_{metric}",
|
90 |
negate * self.metrics[metric](output, y),
|
91 |
on_step=False,
|
92 |
on_epoch=True,
|
|
|
14 |
from remfx.callbacks import log_wandb_audio_batch
|
15 |
from remfx import effects
|
16 |
import asteroid
|
17 |
+
import random
|
18 |
|
19 |
ALL_EFFECTS = effects.Pedalboard_Effects
|
20 |
|
21 |
|
22 |
class RemFXChainInference(pl.LightningModule):
|
23 |
+
def __init__(self, models, sample_rate, num_bins, effect_order):
|
24 |
super().__init__()
|
25 |
self.model = models
|
26 |
self.mrstftloss = MultiResolutionSTFTLoss(
|
|
|
34 |
"FAD": FADLoss(sample_rate=sample_rate),
|
35 |
}
|
36 |
)
|
37 |
+
self.sample_rate = sample_rate
|
38 |
+
self.effect_order = effect_order
|
39 |
|
40 |
+
def forward(self, batch, order=None):
|
41 |
x, y, _, rem_fx_labels = batch
|
42 |
# Use chain of effects defined in config
|
43 |
+
if order:
|
44 |
+
effects_order = order
|
45 |
+
else:
|
46 |
+
effects_order = self.effect_order
|
47 |
effects = [
|
48 |
[ALL_EFFECTS[i] for i, effect in enumerate(effect_label) if effect == 1.0]
|
49 |
for effect_label in rem_fx_labels
|
50 |
]
|
51 |
output = []
|
52 |
with torch.no_grad():
|
53 |
+
for i, (elem, effects_list) in enumerate(zip(x, effects)):
|
54 |
elem = elem.unsqueeze(0) # Add batch dim
|
55 |
+
effect_chain = [
|
56 |
+
effects_order.index(effect.__name__) for effect in effects_list
|
57 |
+
]
|
58 |
+
# log_wandb_audio_batch(
|
59 |
+
# logger=self.logger,
|
60 |
+
# id=f"{i}_Before",
|
61 |
+
# samples=elem.cpu(),
|
62 |
+
# sampling_rate=self.sample_rate,
|
63 |
+
# caption=effect_chain,
|
64 |
+
# )
|
65 |
+
effect_chain
|
66 |
for effect in effect_chain:
|
67 |
+
# Sample the model
|
68 |
+
elem = self.model[effect].model.sample(elem)
|
69 |
+
# log_wandb_audio_batch(
|
70 |
+
# logger=self.logger,
|
71 |
+
# id=f"{i}_{effect}",
|
72 |
+
# samples=elem.cpu(),
|
73 |
+
# sampling_rate=self.sample_rate,
|
74 |
+
# caption=effect_chain,
|
75 |
+
# )
|
|
|
76 |
output.append(elem.squeeze(0))
|
77 |
output = torch.stack(output)
|
78 |
|
|
|
81 |
|
82 |
def test_step(self, batch, batch_idx):
|
83 |
x, y, _, _ = batch # x, y = (B, C, T), (B, C, T)
|
84 |
+
# Random order
|
85 |
+
order = random.shuffle(self.effect_order)
|
86 |
+
loss, output = self.forward(batch, order=order)
|
87 |
# Crop target to match output
|
88 |
if output.shape[-1] < y.shape[-1]:
|
89 |
y = causal_crop(y, output.shape[-1])
|
|
|
97 |
else:
|
98 |
negate = 1
|
99 |
self.log(
|
100 |
+
f"test_{metric}_" + "".join(order),
|
101 |
negate * self.metrics[metric](output, y),
|
102 |
on_step=False,
|
103 |
on_epoch=True,
|
scripts/chain_inference.py
CHANGED
@@ -22,7 +22,7 @@ def main(cfg: DictConfig):
|
|
22 |
model = hydra.utils.instantiate(cfg.model, _convert_="partial")
|
23 |
state_dict = torch.load(ckpt_path)["state_dict"]
|
24 |
model.load_state_dict(state_dict)
|
25 |
-
model.to(
|
26 |
models[effect] = model
|
27 |
|
28 |
callbacks = []
|
@@ -48,7 +48,10 @@ def main(cfg: DictConfig):
|
|
48 |
)
|
49 |
|
50 |
inference_model = RemFXChainInference(
|
51 |
-
models,
|
|
|
|
|
|
|
52 |
)
|
53 |
trainer.test(model=inference_model, datamodule=datamodule)
|
54 |
|
|
|
22 |
model = hydra.utils.instantiate(cfg.model, _convert_="partial")
|
23 |
state_dict = torch.load(ckpt_path)["state_dict"]
|
24 |
model.load_state_dict(state_dict)
|
25 |
+
model.to("cuda" if torch.cuda.is_available() else "cpu")
|
26 |
models[effect] = model
|
27 |
|
28 |
callbacks = []
|
|
|
48 |
)
|
49 |
|
50 |
inference_model = RemFXChainInference(
|
51 |
+
models,
|
52 |
+
sample_rate=cfg.sample_rate,
|
53 |
+
num_bins=cfg.num_bins,
|
54 |
+
order=["Distortion", "Compressor", "Reverb", "Chorus", "Delay"],
|
55 |
)
|
56 |
trainer.test(model=inference_model, datamodule=datamodule)
|
57 |
|