[update] add ds transformation

audioenhancer/constants.py

@@ -4,10 +4,10 @@
 
 SAMPLING_RATE = 44100
 MAX_AUDIO_LENGTH = 10
-BATCH_SIZE = 8
+BATCH_SIZE = 2
 EPOCH = 1
 LOGGING_STEPS = 10
-GRADIENT_ACCUMULATION_STEPS = 1
+GRADIENT_ACCUMULATION_STEPS = 3
 SAVE_STEPS = 100
 EVAL_STEPS = 100
 

audioenhancer/dataset/loader.py

@@ -8,8 +8,8 @@
 
 import dac
 import torch
-import torchaudio
 from audiotools import AudioSignal
+from audiotools import transforms as tfm
 from torch.utils.data import Dataset
 
 
@@ -23,6 +23,18 @@ def __init__(
         mono: bool = True,
         input_freq: int = 16000,
         output_freq: int = 16000,
+        transform: list = [
+            tfm.CorruptPhase,
+            tfm.FrequencyNoise,
+            tfm.HighPass,
+            tfm.LowPass,
+            tfm.MuLawQuantization,
+            tfm.NoiseFloor,
+            tfm.Quantization,
+            tfm.Smoothing,
+            tfm.TimeNoise,
+        ],
+        overall_prob: float = 0.5,
     ):
         """Initializes the dataset.
 
@@ -32,6 +44,8 @@ def __init__(
             mono (bool): Whether to load the audio as mono.
             input_freq (int): The input frequency of the audio.
             output_freq (int): The output frequency of the audio.
+            transform (list): The list of transforms to apply to the audio.
+            overall_prob (float): The overall probability of applying the transforms.
         """
 
         super().__init__()
@@ -58,6 +72,9 @@ def __init__(
         self.autoencoder.eval()
         self.autoencoder.requires_grad_(False)
 
+        prob = overall_prob / len(transform)
+        self._transform = tfm.Compose([trsfm(prob=prob) for trsfm in transform])
+
     def __len__(self) -> int:
         """Returns the number of waveforms in the dataset.
 
@@ -91,6 +108,9 @@ def __getitem__(self, index: int) -> tuple:
         base_waveform = base_waveform.resample(self.autoencoder.sample_rate)
         compressed_waveform = compressed_waveform.resample(self.autoencoder.sample_rate)
 
+        kwargs = self._transform.instantiate(signal=compressed_waveform.clone())
+        compressed_waveform = self._transform(compressed_waveform.clone(), **kwargs)
+
         compressed_waveform = compressed_waveform[:, :, : self._pad_length_input]
         base_waveform = base_waveform[:, :, : self._pad_length_output]
 

audioenhancer/model/audio_ae/model.py

@@ -165,14 +165,15 @@
     dim_out=1024,
     max_seq_len=0,
     attn_layers=Encoder(
-        dim=512,
-        depth=12,
+        dim=1024,
+        depth=16,
         heads=8,
         attn_flash=True,
         cross_attend=False,
         zero_init_branch_output=True,
         rotary_pos_emb=True,
         ff_swish=True,
         ff_glu=True,
+        use_scalenorm=True,
     ),
 )

scripts/train.py

@@ -157,12 +157,6 @@ def eval_model(model, test_loader):
         y = batch[1].to(device, dtype=dtype)
         c, d = x.shape[1], x.shape[2]
 
-        # normalize x over the last dimension
-
-        mean_x = x.mean(dim=-1, keepdim=True)
-        std_x = x.std(dim=-1, keepdim=True)
-        x = (x - mean_x) / std_x
-
         # rearrange x and y
         x = rearrange(x, "b c d t -> b (t c) d")