Revert to initial stoi indexing (#12)

* Revert to initial stoi indexing

* Update to Python 3.8 in CI

.github/workflows/ci.yml

@@ -10,10 +10,10 @@ jobs:
     timeout-minutes: 10
     steps:
     - uses: actions/checkout@v2
-    - name: Set up Python 3.6
+    - name: Set up Python 3.8
       uses: actions/setup-python@v2
       with:
-        python-version: 3.6
+        python-version: 3.8
 
     - name: Install python dependencies
       run: |

torch_stoi/stoi.py

@@ -124,6 +124,9 @@ def forward(self, est_targets: torch.Tensor,
                 targets, est_targets, self.dyn_range, self.win, self.win_len,
                 self.win_len//2
             )
+            # Remove the last mask frame to replicate pystoi behavior
+            mask, _ = mask.sort(-1, descending=True)
+            mask = mask[..., 1:]
 
         # Here comes the real computation, take STFT
         x_spec = self.stft(targets, self.win, self.nfft, overlap=2)
@@ -256,9 +259,7 @@ def remove_silent_frames(x, y, dyn_range, window, framelen, hop):
         x_frames = x_frames.permute(0, 2, 1)
         y_frames = y_frames.permute(0, 2, 1)
 
-        mask, _ = mask.long().sort(-1, descending=True)
-
-        return x_sil, y_sil, mask
+        return x_sil, y_sil, mask.long()
 
     @staticmethod
     def stft(x, win, fft_size, overlap=4):
@@ -272,7 +273,7 @@ def stft(x, win, fft_size, overlap=4):
         win_len = win.shape[0]
         hop = int(win_len / overlap)
         frames = unfold(x[:, None, None, :], kernel_size=(1, win_len),
-                        stride=(1, hop))
+                        stride=(1, hop))[..., :-1]
         return torch.fft.rfft(frames*win[:, None], n=fft_size, dim=1)
 
     @staticmethod