Merge pull request #72 from sp-nitech/magic_intpl

Bug fix of magic_intpl

diffsptk/modules/ap.py

@@ -246,7 +246,7 @@ def forward(self, x, f0):
 
             H = torch.cat((H_alpha, H_beta), dim=-1)  # (B, N, J, 6)
             w = self.window[i, : self.segment_length[i]]  # (J,)
-            Hw = H.mT * w  # (B, N, 6, J)
+            Hw = H.transpose(-2, -1) * w  # (B, N, 6, J)
             R = torch.matmul(Hw, H)  # (B, N, 6, 6)
 
             index_gamma = origin.unsqueeze(-1) + j[..., 1:-1]  # (B, N, J)

diffsptk/modules/excite.py

@@ -91,10 +91,10 @@ def _forward(p, frame_period, voiced_region, unvoiced_region):
 
         # Interpolate pitch.
         if p.dim() != 1:
-            p = p.mT
+            p = p.transpose(-2, -1)
         p = LinearInterpolation._func(p, frame_period)
         if p.dim() != 1:
-            p = p.mT
+            p = p.transpose(-2, -1)
         p *= mask
 
         # Compute phase.

diffsptk/modules/gmm.py

@@ -318,7 +318,7 @@ def forward(self, x):
                     y = posterior.sum(dim=0)
                     nu = px / y.view(-1, 1)
                     nm = torch.matmul(nu.unsqueeze(-1), self.mu.unsqueeze(-2))
-                    mn = nm.mT
+                    mn = nm.transpose(-2, -1)
                     a = pxx - y.view(-1, 1, 1) * (nm + mn - mm)
                     b = xi.view(-1, 1, 1) * self.ubm_sigma
                     diff = self.ubm_mu - self.mu

diffsptk/modules/linear_intpl.py

@@ -79,15 +79,15 @@ def _forward(x, upsampling_factor):
         assert x.dim() == 3, "Input must be 3D tensor"
         B, T, D = x.shape
 
-        x = x.mT.contiguous()  # (B, D, T)
+        x = x.transpose(-2, -1).contiguous()  # (B, D, T)
         x = replicate1(x, left=False)
         x = F.interpolate(
             x,
             size=T * upsampling_factor + 1,
             mode="linear",
             align_corners=True,
         )[..., :-1]  # Remove the padded value.
-        y = x.mT.reshape(B, -1, D)
+        y = x.transpose(-2, -1).reshape(B, -1, D)
 
         if d == 1:
             y = y.view(-1)

diffsptk/modules/magic_intpl.py

@@ -114,7 +114,7 @@ def compute_lerp_inputs(x, magic_number):
                 if uniques[0]:
                     w[..., : counts[0]] = 0
                 w = torch.cumsum(w, dim=-1)
-                w = w - torch.cumsum(w * ~is_magic_number[i], dim=-1)
+                w = w - torch.cummax(w * ~is_magic_number[i], dim=-1)[0]
                 if uniques[0]:
                     w[..., : counts[0]] = 1
                 if uniques[-1]:
@@ -141,10 +141,10 @@ def compute_lerp_inputs(x, magic_number):
             weights = torch.stack(weights)
             return starts, ends, weights
 
-        x = x.mT.reshape(B * D, T)
+        x = x.transpose(-2, -1).reshape(B * D, T)
         starts, ends, weights = compute_lerp_inputs(x, magic_number)
         y = torch.lerp(starts, ends, weights)
-        y = y.reshape(B, D, T).mT
+        y = y.reshape(B, D, T).transpose(-2, -1)
 
         if d == 1:
             y = y.view(-1)

diffsptk/modules/pitch.py

@@ -263,7 +263,7 @@ def calc_embed(self, x):
 
     def calc_pitch(self, x):
         # Compute pitch probabilities.
-        prob = self.calc_prob(x).mT
+        prob = self.calc_prob(x).transpose(-2, -1)
 
         # Decode pitch probabilities.
         pitch, periodicity = self.torchcrepe.postprocess(

diffsptk/modules/rlevdur.py

@@ -91,7 +91,7 @@ def _forward(a, eye):
         E = torch.stack(E[::-1], dim=-1)
 
         V = torch.linalg.solve_triangular(U, eye, upper=True, unitriangular=True)
-        r = torch.matmul(V[..., :1].mT * E, V).squeeze(-2)
+        r = torch.matmul(V[..., :1].transpose(-2, -1) * E, V).squeeze(-2)
         return r
 
     @staticmethod

diffsptk/modules/unframe.py

@@ -127,7 +127,7 @@ def fold(x):
             return x
 
         w = window.repeat(1, 1, N)
-        x = y.mT
+        x = y.transpose(-2, -1)
 
         if d == 2:
             x = x.unsqueeze(0)

tests/test_magic_intpl.py

@@ -23,7 +23,16 @@
 
 @pytest.mark.parametrize("device", ["cpu", "cuda"])
 @pytest.mark.parametrize("module", [False, True])
-def test_compatibility(device, module, N=10, L=2, magic_number=0):
+@pytest.mark.parametrize(
+    "data",
+    [
+        "0 9 0 0 0 0 2 1 0 0 4 5 0 0",
+        "1 9 0 0 0 0 2 1 0 0 4 5 0 7",
+        "1 2 3 4 5 6",
+    ],
+)
+@pytest.mark.parametrize("L", [1, 2])
+def test_compatibility(device, module, data, L, N=10, magic_number=0):
     magic_intpl = U.choice(
         module,
         diffsptk.MagicNumberInterpolation,
@@ -36,14 +45,13 @@ def test_compatibility(device, module, N=10, L=2, magic_number=0):
         device,
         magic_intpl,
         [],
-        "echo 0 9 0 0 0 0 2 1 0 0 4 5 0 0 | x2x +ad",
+        f"echo {data} | x2x +ad",
         f"magic_intpl -l {L} -magic {magic_number}",
         [],
         dx=L,
         dy=L,
     )
 
-    U.check_differentiability(device, magic_intpl, [N, L])
     U.check_differentiability(device, [magic_intpl, F.dropout], [N, L])