fixing issues with residual dancer

models/config/train.yaml

@@ -1,4 +1,4 @@
-training_fn: wav2vec2.train_huggingface
+training_fn: residual.train_residual_dancer
 device: cuda
 seed: 42
 dance_ids: &dance_ids
@@ -19,7 +19,7 @@ dance_ids: &dance_ids
 data_module:
   batch_size: 64
   num_workers: 5
-  test_proportion: 0.2
+  test_proportion: 0.15
 
 datasets:
   preprocessing.dataset.Music4DanceDataset:

models/residual.py

@@ -25,7 +25,6 @@ class ResidualDancer(nn.Module):
         self.n_channels = n_channels
         self.n_classes = n_classes
 
-        # Spectrogram
         self.spec_bn = nn.BatchNorm2d(1)
 
         # CNN
@@ -111,6 +110,7 @@ def train_residual_dancer(config: dict):
     TARGET_CLASSES = config["dance_ids"]
     DEVICE = config["device"]
     SEED = config["seed"]
+    torch.set_float32_matmul_precision('medium')
     pl.seed_everything(SEED, workers=True)
     feature_extractor = SpectrogramTrainingPipeline(**config["feature_extractor"])
     dataset = get_datasets(config["datasets"], feature_extractor)

preprocessing/dataset.py

@@ -73,7 +73,13 @@ class SongDataset(Dataset):
 
         waveform = self._waveform_from_index(idx)
         dance_labels = self._label_from_index(idx)
-        return waveform, dance_labels
+
+        if self._validate_output(waveform, dance_labels):
+            return waveform, dance_labels
+        else:
+            # WARNING: Could cause train/test split leak
+            return self[idx-1]
+
 
     def _idx2audio_idx(self, idx: int) -> int:
         return self._get_audio_loc_from_idx(idx)[0]

preprocessing/pipelines.py

@@ -115,7 +115,4 @@ class AudioToSpectrogram:
     def __call__(self, waveform: torch.Tensor) -> torch.Tensor:
         spectrogram = self.spec(waveform)
         spectrogram = self.to_db(spectrogram)
-
-        # Normalize
-        spectrogram = (spectrogram - spectrogram.mean()) / (2 * spectrogram.std())
         return spectrogram