Reproduce results with latest librosa and torch

datasets/speech_commands_dataset.py

@@ -80,7 +80,7 @@ def __init__(self, folder, transform=None, sample_rate=16000, sample_length=1):
         samples = []
         for f in audio_files:
             path = os.path.join(folder, f)
-            s, sr = librosa.load(path, sample_rate)
+            s, sr = librosa.load(path, sr=sample_rate)
             samples.append(s)
 
         samples = np.hstack(samples)

requirements.txt

@@ -0,0 +1,8 @@
+librosa==0.10.1
+matplotlib==3.8.2
+tensorboardX==2.6.2.2
+torch==2.1.2
+torchaudio==2.1.2
+torchvision==0.16.2
+tqdm==4.66.1
+torchnet==0.0.4

test_speech_commands.py

@@ -89,12 +89,12 @@ def test():
         if args.multi_crop:
             inputs = multi_crop(inputs)
 
-        inputs = Variable(inputs, volatile = True)
+        inputs = Variable(inputs, requires_grad=False)
         targets = Variable(targets, requires_grad=False)
 
         if use_gpu:
             inputs = inputs.cuda()
-            targets = targets.cuda(async=True)
+            targets = targets.cuda()
 
         # forward
         outputs = model(inputs)
@@ -111,7 +111,7 @@ def test():
         pred = outputs.data.max(1, keepdim=True)[1]
         correct += pred.eq(targets.data.view_as(pred)).sum()
         total += targets.size(0)
-        confusion_matrix.add(pred, targets.data)
+        confusion_matrix.add(pred.squeeze(), targets.data)
 
         filenames = batch['path']
         for j in range(len(pred)):

train_speech_commands.py

@@ -152,7 +152,7 @@ def train(epoch):
 
         if use_gpu:
             inputs = inputs.cuda()
-            targets = targets.cuda(async=True)
+            targets = targets.cuda()
 
         # forward/backward
         outputs = model(inputs)
@@ -167,15 +167,15 @@ def train(epoch):
         # statistics
         it += 1
         global_step += 1
-        running_loss += loss.data[0]
+        running_loss += loss.item()
         pred = outputs.data.max(1, keepdim=True)[1]
         if args.mixup:
             targets = batch['target']
-            targets = Variable(targets, requires_grad=False).cuda(async=True)
+            targets = Variable(targets, requires_grad=False).cuda()
         correct += pred.eq(targets.data.view_as(pred)).sum()
         total += targets.size(0)
 
-        writer.add_scalar('%s/loss' % phase, loss.data[0], global_step)
+        writer.add_scalar('%s/loss' % phase, loss.item(), global_step)
 
         # update the progress bar
         pbar.set_postfix({
@@ -210,7 +210,7 @@ def valid(epoch):
 
         if use_gpu:
             inputs = inputs.cuda()
-            targets = targets.cuda(async=True)
+            targets = targets.cuda()
 
         # forward
         outputs = model(inputs)
@@ -219,12 +219,12 @@ def valid(epoch):
         # statistics
         it += 1
         global_step += 1
-        running_loss += loss.data[0]
+        running_loss += loss.item()
         pred = outputs.data.max(1, keepdim=True)[1]
         correct += pred.eq(targets.data.view_as(pred)).sum()
         total += targets.size(0)
 
-        writer.add_scalar('%s/loss' % phase, loss.data[0], global_step)
+        writer.add_scalar('%s/loss' % phase, loss.item(), global_step)
 
         # update the progress bar
         pbar.set_postfix({

transforms/transforms_stft.py

@@ -41,7 +41,7 @@ def __call__(self, data):
         sample_rate = data['sample_rate']
         hop_length = data['hop_length']
         scale = random.uniform(-self.max_scale, self.max_scale)
-        stft_stretch = librosa.core.phase_vocoder(stft, 1+scale, hop_length=hop_length)
+        stft_stretch = librosa.core.phase_vocoder(stft, rate=1+scale, hop_length=hop_length)
         data['stft'] = stft_stretch
         return data
 
@@ -108,7 +108,7 @@ def __call__(self, data):
         stft = data['stft']
         sample_rate = data['sample_rate']
         n_fft = data['n_fft']
-        mel_basis = librosa.filters.mel(sample_rate, n_fft, self.n_mels)
+        mel_basis = librosa.filters.mel(sr=sample_rate, n_fft=n_fft, n_mels=self.n_mels)
         s = np.dot(mel_basis, np.abs(stft)**2.0)
         data['mel_spectrogram'] = librosa.power_to_db(s, ref=np.max)
         return data

transforms/transforms_wav.py

@@ -22,7 +22,7 @@ def __init__(self, sample_rate=16000):
     def __call__(self, data):
         path = data['path']
         if path:
-            samples, sample_rate = librosa.load(path, self.sample_rate)
+            samples, sample_rate = librosa.load(path, sr=self.sample_rate)
         else:
             # silence
             sample_rate = self.sample_rate
@@ -137,7 +137,7 @@ def __init__(self, n_mels=32):
     def __call__(self, data):
         samples = data['samples']
         sample_rate = data['sample_rate']
-        s = librosa.feature.melspectrogram(samples, sr=sample_rate, n_mels=self.n_mels)
+        s = librosa.feature.melspectrogram(y=samples, sr=sample_rate, n_mels=self.n_mels)
         data['mel_spectrogram'] = librosa.power_to_db(s, ref=np.max)
         return data