Dataset work for audio quality processor

codes/data/__init__.py

@@ -75,6 +75,8 @@ def create_dataset(dataset_opt, return_collate=False):
         collate = C(dataset_opt)
     elif mode == 'unsupervised_audio':
         from data.audio.unsupervised_audio_dataset import UnsupervisedAudioDataset as D
+    elif mode == 'unsupervised_audio_with_noise':
+        from data.audio.audio_with_noise_dataset import AudioWithNoiseDataset as D
     else:
         raise NotImplementedError('Dataset [{:s}] is not recognized.'.format(mode))
     dataset = D(dataset_opt)

codes/data/audio/audio_with_noise_dataset.py

@@ -109,7 +109,7 @@ class AudioWithNoiseDataset(Dataset):
                 clip = clip + aug
                 clip.clip_(-1, 1)
         except:
-            print("Exception encountered processing {item}, re-trying because this is often just a failed aug.")
+            print(f"Exception encountered processing {item}, re-trying because this is often just a failed aug.")
             return self[item]
 
         out['clip'] = clip

codes/data/audio/unsupervised_audio_dataset.py

@@ -2,6 +2,7 @@ import os
 import pathlib
 import random
 import sys
+from warnings import warn
 
 import torch
 import torch.utils.data
@@ -32,8 +33,8 @@ def load_audio(audiopath, sampling_rate):
             audio = audio[:, 0]
 
     if lsr != sampling_rate:
-        if lsr < sampling_rate:
+        #if lsr < sampling_rate:
-            print(f'{audiopath} has a sample rate of {sampling_rate} which is lower than the requested sample rate of {sampling_rate}. This is not a good idea.')
+        #    warn(f'{audiopath} has a sample rate of {sampling_rate} which is lower than the requested sample rate of {sampling_rate}. This is not a good idea.')
         audio = torch.nn.functional.interpolate(audio.unsqueeze(0).unsqueeze(1), scale_factor=sampling_rate/lsr, mode='nearest', recompute_scale_factor=False).squeeze()
 
     # Check some assumptions about audio range. This should be automatically fixed in load_wav_to_torch, but might not be in some edge cases, where we should squawk.

codes/data/util.py

@@ -578,6 +578,20 @@ def imresize_np(img, scale, antialiasing=True):
     return out_2.numpy()
 
 
+def load_paths_from_cache(paths, cache_path):
+    if not isinstance(paths, list):
+        paths = [paths]
+    if os.path.exists(cache_path):
+        output = torch.load(cache_path)
+    else:
+        print(f"Building cache for contents of {paths}..")
+        output = []
+        for p in paths:
+            output.extend(find_files_of_type('img', p, qualifier=is_audio_file)[0])
+        torch.save(output, cache_path)
+    return output
+
+
 if __name__ == '__main__':
     # test imresize function
     # read images

codes/train.py

@@ -284,7 +284,7 @@ class Trainer:
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_dvae_audio_clips_with_quantizer_compression.yml')
+    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_noisy_audio_clips_classifier.yml')
     parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher')
     parser.add_argument('--local_rank', type=int, default=0)
     args = parser.parse_args()

codes/trainer/injectors/base_injectors.py

@@ -539,8 +539,7 @@ class MelSpectrogramInjector(Injector):
         from munch import munchify
         from models.tacotron2 import hparams
         hp = munchify(hparams.create_hparams())  # Just use the default tacotron values for the MEL spectrogram. Noone uses anything else anyway.
-        self.stft = TacotronSTFT(hp.filter_length, hp.hop_length, hp.win_length,
+        self.stft = TacotronSTFT(hp.filter_length, hp.hop_length, hp.win_length, hp.n_mel_channels, hp.sampling_rate, hp.mel_fmin, hp.mel_fmax)
-            hp.n_mel_channels, hp.sampling_rate, hp.mel_fmin, hp.mel_fmax)
 
     def forward(self, state):
         inp = state[self.input]