Merge pull request #668 from PaddlePaddle/feat

audio feature

third_party/nnAudio/.gitignore

+build
+dist
+*.egg-info/

third_party/nnAudio/nnAudio/Spectrogram.py

@@ -165,9 +165,13 @@ class STFT(torch.nn.Module):
 #         self.kernel_cos = torch.nn.Parameter(self.kernel_cos, requires_grad=self.trainable)
 
         # Applying window functions to the Fourier kernels
+        if window:
             window_mask = torch.tensor(window_mask)
             wsin = kernel_sin * window_mask
             wcos = kernel_cos * window_mask
+        else:
+            wsin = kernel_sin
+            wcos = kernel_cos
         
         if self.trainable==False:
             self.register_buffer('wsin', wsin)
@@ -179,7 +183,6 @@ class STFT(torch.nn.Module):
             self.register_parameter('wsin', wsin)
             self.register_parameter('wcos', wcos)         
         
-
         # Prepare the shape of window mask so that it can be used later in inverse
         self.register_buffer('window_mask', window_mask.unsqueeze(0).unsqueeze(-1))
         

third_party/nnAudio/setup.py

@@ -2,9 +2,6 @@ import setuptools
 import codecs
 import os.path
 
-with open("README.md", "r") as fh:
-    long_description = fh.read()
-
 def read(rel_path):
     here = os.path.abspath(os.path.dirname(__file__))
     with codecs.open(os.path.join(here, rel_path), 'r') as fp:
@@ -24,7 +21,7 @@ setuptools.setup(
     author="KinWaiCheuk",
     author_email="u3500684@connect.hku.hk",
     description="A fast GPU audio processing toolbox with 1D convolutional neural network",
-    long_description=long_description,
+    long_description='',
     long_description_content_type="text/markdown",
     url="https://github.com/KinWaiCheuk/nnAudio",
     packages=setuptools.find_packages(),