Tune effects, replace models with asteroid, remove silence

cfg/effects/all.yaml

@@ -4,13 +4,19 @@ effects:
   chorus:
       _target_: remfx.effects.RandomPedalboardChorus
       sample_rate: ${sample_rate}
+      min_rate_hz: 0.25
+      max_rate_hz: 1.5
+      min_feedback: 0.1
+      max_feedback: 0.4
       min_depth: 0.2
-      min_mix: 0.3
+      max_depth: 0.6
+      min_mix: 0.15
+      max_mix: 0.4
   distortion:
       _target_: remfx.effects.RandomPedalboardDistortion
       sample_rate: ${sample_rate}
-      min_drive_db: 10
-      max_drive_db: 50
+      min_drive_db: 7
+      max_drive_db: 25
   compressor:
       _target_: remfx.effects.RandomPedalboardCompressor
       sample_rate: ${sample_rate}
@@ -26,7 +32,7 @@ effects:
       min_damping: 0.2
       max_damping: 1.0
       min_wet_dry: 0.2
-      max_wet_dry: 0.8
+      max_wet_dry: 0.6
       min_width: 0.2
       max_width: 1.0
   delay:
@@ -35,6 +41,6 @@ effects:
       min_delay_seconds: 0.1
       max_delay_sconds: 1.0
       min_feedback: 0.05
-      max_feedback: 0.6
-      min_mix: 0.2
-      max_mix: 0.7
+      max_feedback: 0.3
+      min_mix: 0.1
+      max_mix: 0.35

cfg/exp/dist.yaml

@@ -0,0 +1,29 @@
+# @package _global_
+defaults:
+  - override /model: umx
+  - override /effects: all
+seed: 12345
+sample_rate: 48000
+chunk_size: 262144 # 5.5s
+logs_dir: "./logs"
+render_files: True
+render_root: "./data"
+accelerator: "gpu"
+log_audio: True
+# Effects
+max_kept_effects: 5
+max_removed_effects: -1
+shuffle_kept_effects: True
+shuffle_removed_effects: False
+num_classes: 5
+effects_to_use:
+  - compressor
+  - distortion
+  - reverb
+  - chorus
+  - delay
+effects_to_remove:
+  - distortion
+datamodule:
+  batch_size: 16
+  num_workers: 8

cfg/model/dcunet.yaml

@@ -9,16 +9,8 @@ model:
   sample_rate: ${sample_rate}
   network:
     _target_: remfx.models.DCUNetModel
-    spec_dim: 257
-    hidden_dim: 768
-    filter_len: 512
-    hop_len: 64
-    block_layers: 4
-    layers: 4
-    kernel_size: 3
-    refine_layers: 1
-    is_mask: True
-    norm: 'ins'
-    act: 'comp'
+    architecture: "DCUNet-10"
+    stft_kernel_size: 512
+    fix_length_mode: "pad"
     sample_rate: ${sample_rate}
     num_bins: 1025

cfg/model/dptnet.yaml

@@ -9,12 +9,16 @@ model:
   sample_rate: ${sample_rate}
   network:
     _target_: remfx.models.DPTNetModel
-    enc_dim: 256
-    feature_dim: 64
-    hidden_dim: 128
-    layer: 6
-    segment_size: 250
-    nspk: 1
+    n_src: 1
+    in_chan: 1
+    out_chan: 1
+    bn_chan: 128
+    hid_size: 128
+    chunk_size: 100
+    n_repeats: 6
     win_len: 2
+    fb_name: "free"
+    kernel_size: 16
+    n_filters: 1
     sample_rate: ${sample_rate}
     num_bins: 1025

remfx/models.py

@@ -13,6 +13,7 @@ from remfx.dptnet import DPTNet_base
 from remfx.dcunet import RefineSpectrogramUnet
 from remfx.tcn import TCN
 from remfx.utils import causal_crop
+import asteroid
 
 
 class RemFX(pl.LightningModule):
@@ -197,7 +198,7 @@ class DiffusionGenerationModel(nn.Module):
 class DPTNetModel(nn.Module):
     def __init__(self, sample_rate, num_bins, **kwargs):
         super().__init__()
-        self.model = DPTNet_base(**kwargs)
+        self.model = asteroid.models.dptnet.DPTNet(**kwargs)
         self.num_bins = num_bins
         self.mrstftloss = MultiResolutionSTFTLoss(
             n_bins=self.num_bins, sample_rate=sample_rate
@@ -217,7 +218,7 @@ class DPTNetModel(nn.Module):
 class DCUNetModel(nn.Module):
     def __init__(self, sample_rate, num_bins, **kwargs):
         super().__init__()
-        self.model = RefineSpectrogramUnet(**kwargs)
+        self.model = asteroid.models.DCUNet(**kwargs)
         self.mrstftloss = MultiResolutionSTFTLoss(
             n_bins=num_bins, sample_rate=sample_rate
         )
@@ -225,7 +226,7 @@ class DCUNetModel(nn.Module):
 
     def forward(self, batch):
         x, target = batch
-        output = self.model(x.squeeze(1)).unsqueeze(1)  # B x 1 x T
+        output = self.model(x.squeeze(1))  # B x 1 x T
         # Crop target to match output
         if output.shape[-1] < target.shape[-1]:
             target = causal_crop(target, output.shape[-1])
@@ -233,7 +234,7 @@ class DCUNetModel(nn.Module):
         return loss, output
 
     def sample(self, x: Tensor) -> Tensor:
-        output = self.model(x.squeeze(1)).unsqueeze(1)  # B x 1 x T
+        output = self.model(x.squeeze(1))  # B x 1 x T
         return output
 
 

setup.py

@@ -48,6 +48,7 @@ setup(
         "pedalboard",
         "frechet_audio_distance",
         "ordered-set",
+        "asteroid",
     ],
     include_package_data=True,
     license="Apache License 2.0",