Refactored config

config.yaml → cfg/config.yaml

@@ -1,6 +1,8 @@
 defaults:
   - _self_
-  - exp: null
+  - model: null
+  - effects: null
+
 seed: 12345
 train: True
 sample_rate: 48000
@@ -26,12 +28,14 @@ datamodule:
     root: ${oc.env:DATASET_ROOT}
     chunk_size_in_sec: 6
     mode: "train"
+    effect_types: ${effects.train_effects}
   val_dataset:
     _target_: remfx.datasets.VocalSet
     sample_rate: ${sample_rate}
     root: ${oc.env:DATASET_ROOT}
     chunk_size_in_sec: 6
     mode: "val"
+    effect_types: ${effects.val_effects}
   batch_size: 16
   num_workers: 8
   pin_memory: True

cfg/effects/distortion.yaml

@@ -0,0 +1,14 @@
+# @package _global_
+effects:
+  train_effects:
+      Distortion:
+        _target_: remfx.effects.RandomPedalboardDistortion
+        sample_rate: ${sample_rate}
+        min_drive_db: -10
+        max_drive_db: 50
+  val_effects:
+      Distortion:
+        _target_: remfx.effects.RandomPedalboardDistortion
+        sample_rate: ${sample_rate}
+        min_drive_db: 25
+        max_drive_db: 25

cfg/exp/demucs_distortion.yaml

@@ -0,0 +1,4 @@
+# @package _global_
+defaults:
+  - override /model: demucs
+  - override /effects: distortion

cfg/exp/umx_distortion.yaml

@@ -0,0 +1,4 @@
+# @package _global_
+defaults:
+  - override /model: umx
+  - override /effects: distortion

{exp → cfg/model}/demucs.yaml

@@ -13,11 +13,4 @@ model:
     audio_channels: 1
     nfft: 4096
     sample_rate: ${sample_rate}
-datamodule:
-  dataset:
-    effect_types:
-        Distortion:
-          _target_: remfx.effects.RandomPedalboardDistortion
-          sample_rate: ${sample_rate}
-          min_drive_db: -10
-          max_drive_db: 50
+

{exp → cfg/model}/umx.yaml

@@ -14,18 +14,4 @@ model:
     n_channels: 1
     alpha: 0.3
     sample_rate: ${sample_rate}
-datamodule:
-  train_dataset:
-    effect_types:
-        Distortion:
-          _target_: remfx.effects.RandomPedalboardDistortion
-          sample_rate: ${sample_rate}
-          min_drive_db: -10
-          max_drive_db: 50
-  val_dataset:
-    effect_types:
-        Distortion:
-          _target_: remfx.effects.RandomPedalboardDistortion
-          sample_rate: ${sample_rate}
-          min_drive_db: -10
-          max_drive_db: 50
+

remfx/datasets.py

@@ -190,6 +190,9 @@ class VocalSet(Dataset):
         self.chunk_size_in_sec = chunk_size_in_sec
         self.sample_rate = sample_rate
         self.mode = mode
+        import pdb
+
+        pdb.set_trace()
 
         mode_path = self.root / self.mode
         self.files = sorted(list(mode_path.glob("./**/*.wav")))
@@ -225,18 +228,14 @@ class VocalSet(Dataset):
 
         # Add random effect if train
         if self.mode == "train":
-            random_effect_idx = torch.rand(1).item() * len(self.effect_types.keys())
-            effect_name = list(self.effect_types.keys())[int(random_effect_idx)]
-            effect = self.effect_types[effect_name]
-            effected_input = effect(resampled_x)
+            effect_idx = torch.rand(1).item() * len(self.effect_types.keys())
         else:
-            # deterministic static effect for eval
+            # deterministic effect for eval
             effect_idx = idx % len(self.effect_types.keys())
-            effect_name = list(self.effect_types.keys())[effect_idx]
-            effect = deterministic_effects[effect_name]
-            effected_input = torch.from_numpy(
-                effect(resampled_x.numpy(), self.sample_rate)
-            )
+        effect_name = list(self.effect_types.keys())[int(effect_idx)]
+        effect = self.effect_types[effect_name]
+        effected_input = effect(resampled_x)
+
         normalized_input = self.normalize(effected_input)
         normalized_target = self.normalize(resampled_x)
         return (normalized_input, normalized_target, effect_name)

scripts/train.py

@@ -7,12 +7,12 @@ from pytorch_lightning.utilities.model_summary import ModelSummary
 log = utils.get_logger(__name__)
 
 
-@hydra.main(version_base=None, config_path="../", config_name="config.yaml")
+@hydra.main(version_base=None, config_path="../cfg", config_name="config.yaml")
 def main(cfg: DictConfig):
     # Apply seed for reproducibility
     if cfg.seed:
         pl.seed_everything(cfg.seed)
-
+    print(cfg)
     log.info(f"Instantiating datamodule <{cfg.datamodule._target_}>.")
     datamodule = hydra.utils.instantiate(cfg.datamodule, _convert_="partial")
     log.info(f"Instantiating model <{cfg.model._target_}>.")