DltResNet audio extractor improvements

• Automatically resample inputs to 16KHz
• Improve documentation

bob/learn/pytorch/preprocessor/audio/DltResNet.py

@@ -13,6 +13,8 @@ from bob.extension import rc
 from bob.extension.download import download_file
 import os
 from bob.ap import Spectrogram
+from scipy.signal import resample
+
 
 class DltResNetExtractor(Preprocessor):
   """ The class for implementing the feature extraction of DltResNet embeddings.
@@ -25,7 +27,7 @@ class DltResNetExtractor(Preprocessor):
       Use gpu for extracting the embeddings.
 
   """
-  
+
   def __init__(self, model_file=None, num_classes=1211, bn_dim=128, cuda_flag=0, test=False, use_res=False):
     """ Init method
 
@@ -43,11 +45,11 @@ class DltResNetExtractor(Preprocessor):
         Called for unit test (Default: False).
     use_res: bool
         Use ResNet model instead of DltResNet model (Default: False).
-    
+
     """
-        
+
     Preprocessor.__init__(self, min_preprocessed_file_size=bn_dim)
-    
+
     # model
     model_type = "emb"
     if use_res:
@@ -77,9 +79,9 @@ class DltResNetExtractor(Preprocessor):
         model_file = model_path
 
       else:
-        # do nothing (used mainly for unit testing) 
+        # do nothing (used mainly for unit testing)
         pass
-    
+
     if test:
       self.network = torch.nn.DataParallel(self.network)
     else:
@@ -118,8 +120,8 @@ class DltResNetExtractor(Preprocessor):
     Returns
     -------
     feature : 2D :py:class:`numpy.ndarray` (floats)
-      The extracted features as a 1d array of size 128 
-    
+      The extracted features as a 1d array of size 128
+
     """
 
     feats = []
@@ -139,10 +141,10 @@ class DltResNetExtractor(Preprocessor):
             k = specs[i*L:(i+1)*L,:]
         else:
             k = specs[i*L:,:]
-        
+
         X = (k - k.mean(axis=0)) / (k.std(axis=0) + 1e-7)
         batch.append(X)
-        if len(batch) == batch_size or i == embedding_count - 2 or i == embedding_count - 1: 
+        if len(batch) == batch_size or i == embedding_count - 2 or i == embedding_count - 1:
           with torch.no_grad():
             in_batch = numpy.array(batch)
             if cuda_flag == 1:
@@ -152,7 +154,7 @@ class DltResNetExtractor(Preprocessor):
             fX = model(input_var).data.cpu().numpy()
             feats.append(fX)
             batch = []
-  
+
     return numpy.vstack(feats)
 
   def __call__(self, audio, annotations=None):
@@ -161,22 +163,29 @@ class DltResNetExtractor(Preprocessor):
     Parameters
     ----------
     audio : :py:class:`numpy.ndarray` (floats)
-      The audio file to extract the features from. 
+      The audio file to extract the features from.
     annotations : None
       Apply annotations if needed
 
     Returns
     -------
     feature : 2D :py:class:`numpy.ndarray` (floats)
-      The extracted features as a 1d array of size 128 
-    
+      The extracted features as a 1d array of size 128
+
     """
     data = audio[1]
     if data.dtype =='int16':
       data = numpy.cast['float'](data)
     if numpy.max(numpy.abs(data)) < 1:
       data = data * 2**15
+
     rate = audio[0]
+    # resample audio to 16KHz so it can work with the model
+    if rate != 16000:
+      samps = round(len(data) * 16000 / rate)  # Number of samples to resample
+      data = resample(data, samps)
+      rate = 16000
+
     win_length_ms = 25
     win_shift_ms = 10
     normalize_mean = True
@@ -204,4 +213,4 @@ class DltResNetExtractor(Preprocessor):
     """
     f= bob.io.base.HDF5File(data_file)
     feats = f.read("feats")
-    return feats
\ No newline at end of file
+    return feats