Remove numpy support in transforms.py and functional.py

torchaudio/functional.py

-import numpy as np
+import math
 import torch
 
 
@@ -245,15 +245,15 @@ def create_dct(n_mfcc, n_mels, norm):
     outdim = n_mfcc
     dim = n_mels
     # http://en.wikipedia.org/wiki/Discrete_cosine_transform#DCT-II
-    n = np.arange(dim)
-    k = np.arange(outdim)[:, np.newaxis]
-    dct = np.cos(np.pi / dim * (n + 0.5) * k)
+    n = torch.arange(dim, dtype=torch.float32)
+    k = torch.arange(outdim, dtype=torch.float32)[:, None]
+    dct = torch.cos(math.pi / dim * (n + 0.5) * k)
     if norm == 'ortho':
-        dct[0] *= 1.0 / np.sqrt(2)
-        dct *= np.sqrt(2.0 / dim)
+        dct[0] *= 1.0 / math.sqrt(2.0)
+        dct *= math.sqrt(2.0 / dim)
     else:
         dct *= 2
-    return torch.Tensor(dct.T)
+    return dct.t()
 
 
 def MFCC(sig, mel_spect, log_mels, s2db, dct_mat):
@@ -302,7 +302,7 @@ def BLC2CBL(tensor):
 
 
 def mu_law_encoding(x, qc):
-    # type: (Tensor/ndarray, int) -> Tensor/ndarray
+    # type: (Tensor, int) -> Tensor
     """Encode signal based on mu-law companding.  For more info see the
     `Wikipedia Entry <https://en.wikipedia.org/wiki/%CE%9C-law_algorithm>`_
 
@@ -317,10 +317,6 @@ def mu_law_encoding(x, qc):
         Tensor: Input after mu-law companding
     """
     mu = qc - 1.
-    if isinstance(x, np.ndarray):
-        x_mu = np.sign(x) * np.log1p(mu * np.abs(x)) / np.log1p(mu)
-        x_mu = ((x_mu + 1) / 2 * mu + 0.5).astype(int)
-    elif isinstance(x, torch.Tensor):
     if not x.dtype.is_floating_point:
         x = x.to(torch.float)
     mu = torch.tensor(mu, dtype=x.dtype)
@@ -331,7 +327,7 @@ def mu_law_encoding(x, qc):
 
 
 def mu_law_expanding(x_mu, qc):
-    # type: (Tensor/ndarray, int) -> Tensor/ndarray
+    # type: (Tensor, int) -> Tensor
     """Decode mu-law encoded signal.  For more info see the
     `Wikipedia Entry <https://en.wikipedia.org/wiki/%CE%9C-law_algorithm>`_
 
@@ -346,10 +342,6 @@ def mu_law_expanding(x_mu, qc):
         Tensor: Input after decoding
     """
     mu = qc - 1.
-    if isinstance(x_mu, np.ndarray):
-        x = ((x_mu) / mu) * 2 - 1.
-        x = np.sign(x) * (np.exp(np.abs(x) * np.log1p(mu)) - 1.) / mu
-    elif isinstance(x_mu, torch.Tensor):
     if not x_mu.dtype.is_floating_point:
         x_mu = x_mu.to(torch.float)
     mu = torch.tensor(mu, dtype=x_mu.dtype)

torchaudio/transforms.py

 from __future__ import division, print_function
 from warnings import warn
+import math
 import torch
-import numpy as np
 from . import functional as F
 
 
@@ -234,7 +234,7 @@ class SpectrogramToDB(object):
         self.multiplier = 10. if stype == "power" else 20.
         self.amin = 1e-10
         self.ref_value = 1.
-        self.db_multiplier = np.log10(np.maximum(self.amin, self.ref_value))
+        self.db_multiplier = math.log10(max(self.amin, self.ref_value))
 
     def __call__(self, spec):
         # numerically stable implementation from librosa
@@ -403,10 +403,10 @@ class MuLawEncoding(object):
         """
 
         Args:
-            x (FloatTensor/LongTensor or ndarray)
+            x (FloatTensor/LongTensor)
 
         Returns:
-            x_mu (LongTensor or ndarray)
+            x_mu (LongTensor)
 
         """
         return F.mu_law_encoding(x, self.qc)
@@ -434,10 +434,10 @@ class MuLawExpanding(object):
         """
 
         Args:
-            x_mu (FloatTensor/LongTensor or ndarray)
+            x_mu (FloatTensor/LongTensor)
 
         Returns:
-            x (FloatTensor or ndarray)
+            x (FloatTensor)
 
         """
         return F.mu_law_expanding(x_mu, self.qc)