Check argument types in 'checkTypes' (pytorch#1363)

Fixes pytorch#1357

Author: colesbury

test/test_nn.py

@@ -1299,6 +1299,14 @@ def test_assignments(get_list, a, b, c):
  self.assertIn('buf', l.state_dict())
  self.assertIs(l.state_dict()['buf'], buf)
 
+ def test_Conv2d_inconsistent_types(self):
+ inputs = Variable(torch.randn(4, 1, 7, 7).float())
+ weights = Variable(torch.randn(1, 1, 3, 3).double())
+ # inconsistent types should raise an exception
+ self.assertRaises(RuntimeError, lambda: nn.functional.conv2d(inputs, weights))
+ # but it should work with the same type
+ nn.functional.conv2d(inputs.float(), weights.float())
+
  @unittest.skipIf(not TEST_CUDA, 'CUDA not available')
  def test_Conv2d_large_workspace(self):
  # These sizes require huge cuDNN workspaces. Make sure we choose a

torch/csrc/nn/THNN_generic.inc.h

@@ -50,6 +50,9 @@ void checkTypes(bool isCuda, thpp::Type type, ...) {
  if (tensor->isCuda() != isCuda) {
  throw invalid_tensor(isCuda ? "CUDA" : "CPU", tensor->isCuda() ? "CUDA" : "CPU");
  }
+ if (tensor->type() != type) {
+ throw invalid_tensor(thpp::toString(type), thpp::toString(tensor->type()));
+ }
  }
 }