Fix r_order can not be 0 bug.

.gitignore

@@ -1 +1,2 @@
-.idea
+.idea
+__pycache__

main.py

@@ -0,0 +1,30 @@
+import torch
+import torch.nn as nn
+
+from module import FSMNKernel
+
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.fsmn = FSMNKernel(220, 0, 0, 1, 1)
+        self.fc = nn.Linear(220, 3)
+
+    def forward(self, x):
+        return self.fc(nn.functional.relu(self.fsmn(x)))
+
+
+if __name__ == '__main__':
+    B, T, D = 32, 300, 220
+    xx = torch.randn(B, T, D)
+    yy = torch.randint(low=0, high=3, size=(32, 300)).view(-1)
+    net = Net()
+    optim = torch.optim.SGD(net.parameters(), lr=1e-2, momentum=0.8)
+    L = nn.CrossEntropyLoss()
+    for _ in range(100):
+        y = net(xx).view(-1, 3)
+        loss = L(y, yy)
+        print(loss.item())
+        optim.zero_grad()
+        loss.backward()
+        optim.step()

module/__init__.py

@@ -0,0 +1 @@
+from .fsmn_kernel import FSMNKernel

fsmn_kernel.py → module/fsmn_kernel.py

@@ -34,26 +34,21 @@ def forward(self, x):
 
         kernel_out = []
         for frame in range(self.l_order * self.l_stride, frames + self.l_order * self.l_stride):
-            cur_frame = torch.sum(
-                x[:, frame - self.l_order * self.l_stride:frame:self.l_stride] * self.filter[:self.l_order] +
-                x[:, frame:frame + 1] * self.filter[self.l_order:self.l_order + 1] +
-                x[:, frame + 1:frame + 1 + self.r_order * self.r_stride:self.r_stride] * self.filter[self.l_order + 1:],
-                dim=1,
-                keepdim=True
-            )
+            if self.r_order > 0:
+                cur_frame = torch.sum(
+                    x[:, frame - self.l_order * self.l_stride:frame:self.l_stride] * self.filter[:self.l_order] +
+                    x[:, frame:frame + 1] * self.filter[self.l_order:self.l_order + 1] +
+                    x[:, frame + 1:frame + 1 + self.r_order * self.r_stride:self.r_stride] * self.filter[self.l_order + 1:],
+                    dim=1,
+                    keepdim=True
+                )
+            else:
+                cur_frame = torch.sum(
+                    x[:, frame - self.l_order * self.l_stride:frame:self.l_stride] * self.filter[:self.l_order] +
+                    x[:, frame:frame + 1] * self.filter[self.l_order:self.l_order + 1],
+                    dim=1,
+                    keepdim=True
+                )
             kernel_out.append(cur_frame)
         kernel_out = torch.cat(kernel_out, dim=1)
         return kernel_out
-
-
-if __name__ == '__main__':
-    B, T, D = 32, 300, 440
-    xx = torch.randn(B, T, D, requires_grad=True)
-    kernel = FSMNKernel(D, 5, 5, 1, 2)
-    out = kernel(xx)
-    print(kernel)
-    loss = out.mean()
-    loss.backward()
-    for name, param in kernel.named_parameters():
-        print(name, param, param.grad)
-