Spaces:
Runtime error
Runtime error
mattricesound
commited on
Commit
•
568c3f1
1
Parent(s):
836d971
Update to latest classifier inference
Browse files- README.md +18 -24
- cfg/exp/5-5_full_cls.yaml +2 -2
- cfg/exp/5-5_full_cls_dynamic.yaml +1 -1
- remfx/classifier.py +2 -15
- remfx/datasets.py +1 -1
- remfx/effects.py +0 -2
- remfx/models.py +3 -13
- remfx/tcn.py +0 -1
- scripts/test.py +2 -1
- setup.py +9 -0
README.md
CHANGED
@@ -10,14 +10,19 @@ git clone https://github.com/mhrice/RemFx.git
|
|
10 |
cd RemFx
|
11 |
git submodule update --init --recursive
|
12 |
pip install -e . ./umx
|
|
|
13 |
```
|
|
|
14 |
# Usage
|
15 |
This repo can be used for many different tasks. Here are some examples.
|
16 |
## Run RemFX Detect on a single file
|
17 |
First, need to download the checkpoints from [zenodo](https://zenodo.org/record/8179396)
|
18 |
```
|
19 |
scripts/download_checkpoints.sh
|
20 |
-
|
|
|
|
|
|
|
21 |
```
|
22 |
## Download the [General Purpose Audio Effect Removal evaluation datasets](https://zenodo.org/record/8187288)
|
23 |
```
|
@@ -69,6 +74,18 @@ If you have generated the dataset separately (see Generate datasets used in the
|
|
69 |
|
70 |
Also note that the training assumes you have a GPU. To train on CPU, set `accelerator=null` in the config or command-line.
|
71 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
## Evaluate models on the General Purpose Audio Effect Removal evaluation datasets (Table 4 from the paper)
|
73 |
First download the General Purpose Audio Effect Removal evaluation datasets (see above).
|
74 |
To use the pretrained RemFX model, download the checkpoints
|
@@ -148,26 +165,3 @@ Some relevant dataset/training parameters descriptions
|
|
148 |
- `distortion`
|
149 |
- `reverb`
|
150 |
- `delay`
|
151 |
-
|
152 |
-
<!-- # DO WE NEED THIS?
|
153 |
-
## Evaluate RemFXwith a custom directory
|
154 |
-
Assumes directory is structured as
|
155 |
-
- root
|
156 |
-
- clean
|
157 |
-
- file1.wav
|
158 |
-
- file2.wav
|
159 |
-
- file3.wav
|
160 |
-
- effected
|
161 |
-
- file1.wav
|
162 |
-
- file2.wav
|
163 |
-
- file3.wav
|
164 |
-
|
165 |
-
First set the dataset root:
|
166 |
-
```
|
167 |
-
export DATASET_ROOT={path/to/datasets}
|
168 |
-
```
|
169 |
-
|
170 |
-
Then run
|
171 |
-
```
|
172 |
-
python scripts/chain_inference.py +exp=chain_inference_custom
|
173 |
-
``` -->
|
|
|
10 |
cd RemFx
|
11 |
git submodule update --init --recursive
|
12 |
pip install -e . ./umx
|
13 |
+
pip install --no-deps hearbaseline
|
14 |
```
|
15 |
+
Due to incompatabilities with hearbaseline's dependencies (namely numpy/numba) and our other packages, we need to install hearbaseline with no dependencies.
|
16 |
# Usage
|
17 |
This repo can be used for many different tasks. Here are some examples.
|
18 |
## Run RemFX Detect on a single file
|
19 |
First, need to download the checkpoints from [zenodo](https://zenodo.org/record/8179396)
|
20 |
```
|
21 |
scripts/download_checkpoints.sh
|
22 |
+
```
|
23 |
+
Then run the detect script. This repo contains an example file `example.wav` from our test dataset which contains 2 effects (chorus and delay) applied to a guitar.
|
24 |
+
```
|
25 |
+
scripts/remfx_detect.sh example.wav -o dry.wav
|
26 |
```
|
27 |
## Download the [General Purpose Audio Effect Removal evaluation datasets](https://zenodo.org/record/8187288)
|
28 |
```
|
|
|
74 |
|
75 |
Also note that the training assumes you have a GPU. To train on CPU, set `accelerator=null` in the config or command-line.
|
76 |
|
77 |
+
### Logging
|
78 |
+
Default CSV logger
|
79 |
+
To use WANDB logger:
|
80 |
+
|
81 |
+
export WANDB_PROJECT={desired_wandb_project}
|
82 |
+
export WANDB_ENTITY={your_wandb_username}
|
83 |
+
|
84 |
+
## Panns pretrianed
|
85 |
+
```
|
86 |
+
wget https://zenodo.org/record/6332525/files/hear2021-panns_hear.pth
|
87 |
+
```
|
88 |
+
|
89 |
## Evaluate models on the General Purpose Audio Effect Removal evaluation datasets (Table 4 from the paper)
|
90 |
First download the General Purpose Audio Effect Removal evaluation datasets (see above).
|
91 |
To use the pretrained RemFX model, download the checkpoints
|
|
|
165 |
- `distortion`
|
166 |
- `reverb`
|
167 |
- `delay`
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
cfg/exp/5-5_full_cls.yaml
CHANGED
@@ -1,11 +1,11 @@
|
|
1 |
# @package _global_
|
2 |
defaults:
|
3 |
-
- override /model:
|
4 |
- override /effects: all
|
5 |
seed: 12345
|
6 |
sample_rate: 48000
|
7 |
chunk_size: 262144 # 5.5s
|
8 |
-
logs_dir: "
|
9 |
render_files: True
|
10 |
|
11 |
accelerator: "gpu"
|
|
|
1 |
# @package _global_
|
2 |
defaults:
|
3 |
+
- override /model: cls_panns_48k_specaugment
|
4 |
- override /effects: all
|
5 |
seed: 12345
|
6 |
sample_rate: 48000
|
7 |
chunk_size: 262144 # 5.5s
|
8 |
+
logs_dir: "./logs"
|
9 |
render_files: True
|
10 |
|
11 |
accelerator: "gpu"
|
cfg/exp/5-5_full_cls_dynamic.yaml
CHANGED
@@ -5,7 +5,7 @@ defaults:
|
|
5 |
seed: 12345
|
6 |
sample_rate: 48000
|
7 |
chunk_size: 262144 # 5.5s
|
8 |
-
logs_dir: "
|
9 |
render_files: True
|
10 |
|
11 |
accelerator: "gpu"
|
|
|
5 |
seed: 12345
|
6 |
sample_rate: 48000
|
7 |
chunk_size: 262144 # 5.5s
|
8 |
+
logs_dir: "./logs"
|
9 |
render_files: True
|
10 |
|
11 |
accelerator: "gpu"
|
remfx/classifier.py
CHANGED
@@ -171,7 +171,6 @@ class Cnn14(nn.Module):
|
|
171 |
|
172 |
self.fc1 = nn.Linear(2048, 2048, bias=True)
|
173 |
|
174 |
-
# self.fc_audioset = nn.Linear(2048, num_classes, bias=True)
|
175 |
self.heads = torch.nn.ModuleList()
|
176 |
for _ in range(num_classes):
|
177 |
self.heads.append(nn.Linear(2048, 1, bias=True))
|
@@ -190,7 +189,6 @@ class Cnn14(nn.Module):
|
|
190 |
def init_weight(self):
|
191 |
init_bn(self.bn0)
|
192 |
init_layer(self.fc1)
|
193 |
-
# init_layer(self.fc_audioset)
|
194 |
|
195 |
def forward(self, x: torch.Tensor, train: bool = False):
|
196 |
"""
|
@@ -202,20 +200,11 @@ class Cnn14(nn.Module):
|
|
202 |
x = self.melspec(x)
|
203 |
|
204 |
if self.specaugment and train:
|
205 |
-
# import matplotlib.pyplot as plt
|
206 |
-
# fig, axs = plt.subplots(2, 1, sharex=True)
|
207 |
-
# axs[0].imshow(x[0, :, :, :].detach().squeeze().cpu().numpy())
|
208 |
x = self.freq_mask(x)
|
209 |
x = self.time_mask(x)
|
210 |
-
# axs[1].imshow(x[0, :, :, :].detach().squeeze().cpu().numpy())
|
211 |
-
# plt.savefig("spec_augment.png", dpi=300)
|
212 |
-
|
213 |
-
# x = x.permute(0, 2, 1, 3)
|
214 |
-
# x = self.bn0(x)
|
215 |
-
# x = x.permute(0, 2, 1, 3)
|
216 |
|
217 |
# apply standardization
|
218 |
-
x = (x - x.mean(dim=
|
219 |
|
220 |
x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
|
221 |
x = F.dropout(x, p=0.2, training=train)
|
@@ -241,8 +230,6 @@ class Cnn14(nn.Module):
|
|
241 |
for head in self.heads:
|
242 |
outputs.append(torch.sigmoid(head(x)))
|
243 |
|
244 |
-
# clipwise_output = self.fc_audioset(x)
|
245 |
-
|
246 |
return outputs
|
247 |
|
248 |
|
@@ -294,4 +281,4 @@ class ConvBlock(nn.Module):
|
|
294 |
else:
|
295 |
raise Exception("Incorrect argument!")
|
296 |
|
297 |
-
return x
|
|
|
171 |
|
172 |
self.fc1 = nn.Linear(2048, 2048, bias=True)
|
173 |
|
|
|
174 |
self.heads = torch.nn.ModuleList()
|
175 |
for _ in range(num_classes):
|
176 |
self.heads.append(nn.Linear(2048, 1, bias=True))
|
|
|
189 |
def init_weight(self):
|
190 |
init_bn(self.bn0)
|
191 |
init_layer(self.fc1)
|
|
|
192 |
|
193 |
def forward(self, x: torch.Tensor, train: bool = False):
|
194 |
"""
|
|
|
200 |
x = self.melspec(x)
|
201 |
|
202 |
if self.specaugment and train:
|
|
|
|
|
|
|
203 |
x = self.freq_mask(x)
|
204 |
x = self.time_mask(x)
|
|
|
|
|
|
|
|
|
|
|
|
|
205 |
|
206 |
# apply standardization
|
207 |
+
x = (x - x.mean(dim=(2, 3), keepdim=True)) / x.std(dim=(2, 3), keepdim=True)
|
208 |
|
209 |
x = self.conv_block1(x, pool_size=(2, 2), pool_type="avg")
|
210 |
x = F.dropout(x, p=0.2, training=train)
|
|
|
230 |
for head in self.heads:
|
231 |
outputs.append(torch.sigmoid(head(x)))
|
232 |
|
|
|
|
|
233 |
return outputs
|
234 |
|
235 |
|
|
|
281 |
else:
|
282 |
raise Exception("Incorrect argument!")
|
283 |
|
284 |
+
return x
|
remfx/datasets.py
CHANGED
@@ -666,7 +666,7 @@ class EffectDatamodule(pl.LightningDataModule):
|
|
666 |
def test_dataloader(self) -> DataLoader:
|
667 |
return DataLoader(
|
668 |
dataset=self.test_dataset,
|
669 |
-
batch_size=
|
670 |
num_workers=self.num_workers,
|
671 |
pin_memory=self.pin_memory,
|
672 |
shuffle=False,
|
|
|
666 |
def test_dataloader(self) -> DataLoader:
|
667 |
return DataLoader(
|
668 |
dataset=self.test_dataset,
|
669 |
+
batch_size=self.test_batch_size,
|
670 |
num_workers=self.num_workers,
|
671 |
pin_memory=self.pin_memory,
|
672 |
shuffle=False,
|
remfx/effects.py
CHANGED
@@ -84,7 +84,6 @@ def biqaud(
|
|
84 |
a2 = 1 - alpha / A
|
85 |
else:
|
86 |
pass
|
87 |
-
# raise ValueError(f"Invalid filter_type: {filter_type}.")
|
88 |
|
89 |
b = np.array([b0, b1, b2]) / a0
|
90 |
a = np.array([a0, a1, a2]) / a0
|
@@ -291,7 +290,6 @@ class RandomVolumeAutomation(torch.nn.Module):
|
|
291 |
gain_db[samples_filled : samples_filled + segment_samples] = fade
|
292 |
samples_filled = samples_filled + segment_samples
|
293 |
|
294 |
-
# print(gain_db)
|
295 |
x *= 10 ** (gain_db / 20.0)
|
296 |
return x
|
297 |
|
|
|
84 |
a2 = 1 - alpha / A
|
85 |
else:
|
86 |
pass
|
|
|
87 |
|
88 |
b = np.array([b0, b1, b2]) / a0
|
89 |
a = np.array([a0, a1, a2]) / a0
|
|
|
290 |
gain_db[samples_filled : samples_filled + segment_samples] = fade
|
291 |
samples_filled = samples_filled + segment_samples
|
292 |
|
|
|
293 |
x *= 10 ** (gain_db / 20.0)
|
294 |
return x
|
295 |
|
remfx/models.py
CHANGED
@@ -55,12 +55,11 @@ class RemFXChainInference(pl.LightningModule):
|
|
55 |
effects_order = order
|
56 |
else:
|
57 |
effects_order = self.effect_order
|
58 |
-
|
59 |
# Use classifier labels
|
60 |
if self.classifier:
|
61 |
threshold = 0.5
|
62 |
with torch.no_grad():
|
63 |
-
labels = torch.
|
64 |
rem_fx_labels = torch.where(labels > threshold, 1.0, 0.0)
|
65 |
if self.use_all_effect_models:
|
66 |
effects_present = [
|
@@ -253,17 +252,8 @@ class RemFX(pl.LightningModule):
|
|
253 |
prog_bar=True,
|
254 |
sync_dist=True,
|
255 |
)
|
256 |
-
# print(f"Input_{metric}", negate * self.metrics[metric](x, y))
|
257 |
-
# print(f"test_{metric}", negate * self.metrics[metric](output, y))
|
258 |
-
# self.output_str += f"{negate * self.metrics[metric](x, y).item():.4f},{negate * self.metrics[metric](output, y).item():.4f},"
|
259 |
-
# self.output_str += "\n"
|
260 |
return loss
|
261 |
|
262 |
-
def on_test_end(self) -> None:
|
263 |
-
pass
|
264 |
-
# with open("output.csv", "w") as f:
|
265 |
-
# f.write(self.output_str)
|
266 |
-
|
267 |
|
268 |
class OpenUnmixModel(nn.Module):
|
269 |
def __init__(
|
@@ -418,7 +408,6 @@ def mixup(x: torch.Tensor, y: torch.Tensor, alpha: float = 1.0):
|
|
418 |
else:
|
419 |
lam = 1
|
420 |
|
421 |
-
print(lam)
|
422 |
if np.random.rand() > 0.5:
|
423 |
index = torch.randperm(batch_size).to(x.device)
|
424 |
mixed_x = lam * x + (1 - lam) * x[index, :]
|
@@ -429,6 +418,7 @@ def mixup(x: torch.Tensor, y: torch.Tensor, alpha: float = 1.0):
|
|
429 |
|
430 |
return mixed_x, mixed_y, lam
|
431 |
|
|
|
432 |
class FXClassifier(pl.LightningModule):
|
433 |
def __init__(
|
434 |
self,
|
@@ -533,4 +523,4 @@ class FXClassifier(pl.LightningModule):
|
|
533 |
lr=self.lr,
|
534 |
weight_decay=self.lr_weight_decay,
|
535 |
)
|
536 |
-
return optimizer
|
|
|
55 |
effects_order = order
|
56 |
else:
|
57 |
effects_order = self.effect_order
|
|
|
58 |
# Use classifier labels
|
59 |
if self.classifier:
|
60 |
threshold = 0.5
|
61 |
with torch.no_grad():
|
62 |
+
labels = torch.hstack(self.classifier(x))
|
63 |
rem_fx_labels = torch.where(labels > threshold, 1.0, 0.0)
|
64 |
if self.use_all_effect_models:
|
65 |
effects_present = [
|
|
|
252 |
prog_bar=True,
|
253 |
sync_dist=True,
|
254 |
)
|
|
|
|
|
|
|
|
|
255 |
return loss
|
256 |
|
|
|
|
|
|
|
|
|
|
|
257 |
|
258 |
class OpenUnmixModel(nn.Module):
|
259 |
def __init__(
|
|
|
408 |
else:
|
409 |
lam = 1
|
410 |
|
|
|
411 |
if np.random.rand() > 0.5:
|
412 |
index = torch.randperm(batch_size).to(x.device)
|
413 |
mixed_x = lam * x + (1 - lam) * x[index, :]
|
|
|
418 |
|
419 |
return mixed_x, mixed_y, lam
|
420 |
|
421 |
+
|
422 |
class FXClassifier(pl.LightningModule):
|
423 |
def __init__(
|
424 |
self,
|
|
|
523 |
lr=self.lr,
|
524 |
weight_decay=self.lr_weight_decay,
|
525 |
)
|
526 |
+
return optimizer
|
remfx/tcn.py
CHANGED
@@ -91,7 +91,6 @@ class TCN(nn.Module):
|
|
91 |
self.causal = causal
|
92 |
self.estimate_loudness = estimate_loudness
|
93 |
|
94 |
-
print(f"Causal: {self.causal}")
|
95 |
if self.causal:
|
96 |
self.crop_fn = causal_crop
|
97 |
else:
|
|
|
91 |
self.causal = causal
|
92 |
self.estimate_loudness = estimate_loudness
|
93 |
|
|
|
94 |
if self.causal:
|
95 |
self.crop_fn = causal_crop
|
96 |
else:
|
scripts/test.py
CHANGED
@@ -16,7 +16,8 @@ def main(cfg: DictConfig):
|
|
16 |
datamodule = hydra.utils.instantiate(cfg.datamodule, _convert_="partial")
|
17 |
log.info(f"Instantiating model <{cfg.model._target_}>.")
|
18 |
model = hydra.utils.instantiate(cfg.model, _convert_="partial")
|
19 |
-
|
|
|
20 |
"state_dict"
|
21 |
]
|
22 |
model.load_state_dict(state_dict)
|
|
|
16 |
datamodule = hydra.utils.instantiate(cfg.datamodule, _convert_="partial")
|
17 |
log.info(f"Instantiating model <{cfg.model._target_}>.")
|
18 |
model = hydra.utils.instantiate(cfg.model, _convert_="partial")
|
19 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
20 |
+
state_dict = torch.load(cfg.ckpt_path, map_location=device)[
|
21 |
"state_dict"
|
22 |
]
|
23 |
model.load_state_dict(state_dict)
|
setup.py
CHANGED
@@ -44,6 +44,15 @@ setup(
|
|
44 |
"pyloudnorm",
|
45 |
"pedalboard",
|
46 |
"asteroid",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
],
|
48 |
include_package_data=True,
|
49 |
license="Apache License 2.0",
|
|
|
44 |
"pyloudnorm",
|
45 |
"pedalboard",
|
46 |
"asteroid",
|
47 |
+
"librosa",
|
48 |
+
"speechbrain",
|
49 |
+
"torchcrepe",
|
50 |
+
"torchopenl3",
|
51 |
+
"tensorflow",
|
52 |
+
"transformers",
|
53 |
+
"torchmetrics>=1.0",
|
54 |
+
"wav2clip_hear @ git+https://github.com/hohsiangwu/wav2clip-hear.git",
|
55 |
+
"panns_hear @ git+https://github.com/qiuqiangkong/HEAR2021_Challenge_PANNs",
|
56 |
],
|
57 |
include_package_data=True,
|
58 |
license="Apache License 2.0",
|