Add max_removed_effects

README.md

@@ -27,20 +27,21 @@
 - `distortion`
 - `reverb`
 
-## Train CLI Options
+## Train Main CLI Options
 - `max_kept_effects={n}` max number of <b> Kept </b> effects to apply to each file (default: 3)
+- `max_removed_effects={n}` max number of <b> Removed </b> effects to apply to each file (default: 4)
 - `model={model}` architecture to use (see 'Models')
 - `shuffle_kept_effects=True/False` Shuffle kept effects (default: True)
-- `shuffle_removed_effects=True/False` Shuffle Removed effects (default: False)
+- `shuffle_removed_effects=True/False` Shuffle removed effects (default: False)
 - `effects_to_use={effect}` Effects to use (see 'Effects') (default: all in the list)
 - `effects_to_remove={effect}` Effects to remove (see 'Effects') (default: all in the list)
-- `trainer.accelerator='gpu'` : Use GPU (default: None)
-- `trainer.devices={n}` Number of GPUs to use (default: 1)
+- `accelerator=null/gpu` Use GPU (1 device) (default: False)
 - `render_files=True/False` Render files. Disable to skip rendering stage (default: True)
 - `render_root={path/to/dir}`. Root directory to render files to (default: DATASET_ROOT)
 
-Example: `python scripts/train.py model=demucs "effects_to_use=[distortion, reverb]" "effects_to_remove=[distortion]" "max_kept_effects=2" "shuffle_kept_effects=False" "shuffle_removed_effects=True" trainer.accelerator='gpu' trainer.devices=2`
+Example: `python scripts/train.py model=demucs "effects_to_use=[distortion, reverb]" "effects_to_remove=[distortion]" max_kept_effects=2 max_removed_effects=4 shuffle_kept_effects=False shuffle_removed_effects=True accelerator='gpu' render_root='/home/username/datasets/vocalset'`
 
+See `cfg/config.yaml` for more options that can be specified on the command line.
 
 ## Misc.
 By default, files are rendered to `input_dir / processed / {string_of_effects} / {train|val|test}`.

cfg/config.yaml

@@ -11,8 +11,10 @@ chunk_size: 262144 # 5.5s
 logs_dir: "./logs"
 render_files: True
 render_root: "./data"
+accelerator: null
 
 max_kept_effects: 3
+max_removed_effects: 4
 shuffle_kept_effects: True
 shuffle_removed_effects: False
 effects_to_use:
@@ -52,6 +54,7 @@ datamodule:
     effects_to_use: ${effects_to_use}
     effects_to_remove: ${effects_to_remove}
     max_kept_effects: ${max_kept_effects}
+    max_removed_effects: ${max_removed_effects}
     shuffle_kept_effects: ${shuffle_kept_effects}
     shuffle_removed_effects: ${shuffle_removed_effects}
     render_files: ${render_files}
@@ -66,6 +69,7 @@ datamodule:
     effects_to_use: ${effects_to_use}
     effects_to_remove: ${effects_to_remove}
     max_kept_effects: ${max_kept_effects}
+    max_removed_effects: ${max_removed_effects}
     shuffle_kept_effects: ${shuffle_kept_effects}
     shuffle_removed_effects: ${shuffle_removed_effects}
     render_files: ${render_files}
@@ -80,6 +84,7 @@ datamodule:
     effects_to_use: ${effects_to_use}
     effects_to_remove: ${effects_to_remove}
     max_kept_effects: ${max_kept_effects}
+    max_removed_effects: ${max_removed_effects}
     shuffle_kept_effects: ${shuffle_kept_effects}
     shuffle_removed_effects: ${shuffle_removed_effects}
     render_files: ${render_files}
@@ -107,7 +112,7 @@ trainer:
   enable_model_summary: False
   log_every_n_steps: 1 # Logs metrics every N batches
   accumulate_grad_batches: 1
-  accelerator: null
+  accelerator: ${accelerator}
   devices: 1
   gradient_clip_val: 10.0
   max_steps: 50000

remfx/datasets.py

@@ -26,7 +26,8 @@ class VocalSet(Dataset):
         effect_modules: List[Dict[str, torch.nn.Module]] = None,
         effects_to_use: List[str] = None,
         effects_to_remove: List[str] = None,
-        max_kept_effects: int = 1,
+        max_kept_effects: int = 4,
+        max_removed_effects: int = 4,
         shuffle_kept_effects: bool = True,
         shuffle_removed_effects: bool = False,
         render_files: bool = True,
@@ -44,6 +45,7 @@ class VocalSet(Dataset):
         mode_path = self.root / self.mode
         self.files = sorted(list(mode_path.glob("./**/*.wav")))
         self.max_kept_effects = max_kept_effects
+        self.max_removed_effects = max_removed_effects
         self.effects_to_use = effects_to_use
         self.effects_to_remove = effects_to_remove
         self.normalize = effects.LoudnessNormalize(sample_rate, target_lufs_db=-20)
@@ -130,12 +132,12 @@ class VocalSet(Dataset):
                 )
         kept_fx = list(set(self.effects_to_use) - set(self.effects_to_remove))
         kept_str = "randomly" if self.shuffle_kept_effects else "in order"
-        removed_str = "randomly" if self.shuffle_removed_effects else "in order"
         rem_fx = self.effects_to_remove
+        rem_str = "randomly" if self.shuffle_removed_effects else "in order"
         print(
             f"Effect Summary: \n"
             f"Apply effects: {kept_fx} (Up to {self.max_kept_effects}, chosen {kept_str}) -> Dry\n"
-            f"Apply effects: {rem_fx} (All {len(rem_fx)}, chosen {removed_str}) -> Wet\n"
+            f"Apply effects: {rem_fx} (Up to {self.max_removed_effects}, chosen {rem_str}) -> Wet\n"
         )
         return kept_fx
 
@@ -143,25 +145,34 @@ class VocalSet(Dataset):
         labels = []
 
         # Apply Kept Effects
+        # Shuffle effects if specified
         if self.shuffle_kept_effects:
             effect_indices = torch.randperm(len(self.effects_to_keep))
         else:
             effect_indices = torch.arange(len(self.effects_to_keep))
+        # Up to max_kept_effects
+        effect_indices = effect_indices[: self.max_kept_effects]
+        # Index in effect settings
         effect_names_to_apply = [self.effects_to_keep[i] for i in effect_indices]
         effects_to_apply = [self.effects[i] for i in effect_names_to_apply]
+        # Apply
         for effect in effects_to_apply:
             dry = effect(dry)
             labels.append(ALL_EFFECTS.index(type(effect)))
-        print(labels)
 
         # Apply effects_to_remove
+        # Shuffle effects if specified
         wet = torch.clone(dry)
         if self.shuffle_removed_effects:
             effect_indices = torch.randperm(len(self.effects_to_remove))
         else:
             effect_indices = torch.arange(len(self.effects_to_remove))
+        # Up to max_removed_effects
+        effect_indices = effect_indices[: self.max_removed_effects]
+        # Index in effect settings
         effect_names_to_apply = [self.effects_to_remove[i] for i in effect_indices]
         effects_to_apply = [self.effects[i] for i in effect_names_to_apply]
+        # Apply
         for effect in effects_to_apply:
             wet = effect(wet)
             labels.append(ALL_EFFECTS.index(type(effect)))