replace bmm with matmul due to driver issues

Pointnet_Example/model.py

@@ -73,13 +73,13 @@ class PointNet(nn.Module):
         def forward(self, input):
             matrix3x3 = self.input_transform(input)
             # batch matrix multiplication
-            xb = torch.bmm(torch.transpose(input, 1, 2),
+            xb = torch.matmul(torch.transpose(input, 1, 2),
                            matrix3x3).transpose(1, 2)
 
             xb = F.relu(self.bn1(self.conv1(xb)))
 
             matrix64x64 = self.feature_transform(xb)
-            xb = torch.bmm(torch.transpose(xb, 1, 2),
+            xb = torch.matmul(torch.transpose(xb, 1, 2),
                            matrix64x64).transpose(1, 2)
 
             xb = F.relu(self.bn2(self.conv2(xb)))
@@ -104,8 +104,8 @@ def pointnetloss(outputs, labels, m3x3, m64x64, alpha=0.001, device=None):
     # Calculate difference to identity matrix for regularization.
     id3x3 = torch.eye(3, requires_grad=True, device=device).repeat(bs, 1, 1)
     id64x64 = torch.eye(64, requires_grad=True, device=device).repeat(bs, 1, 1)
-    diff3x3 = id3x3 - torch.bmm(m3x3, m3x3.transpose(1, 2))
-    diff64x64 = id64x64 - torch.bmm(m64x64, m64x64.transpose(1, 2))
+    diff3x3 = id3x3 - torch.matmul(m3x3, m3x3.transpose(1, 2))
+    diff64x64 = id64x64 - torch.matmul(m64x64, m64x64.transpose(1, 2))
     # Negative log likelihood criterion is already adapted to batch size.
     return criterion(outputs, labels) + alpha * (torch.norm(diff3x3) + torch.norm(diff64x64)) / float(bs)