Fix gh-2210

theano/tensor/basic.py

@@ -3474,7 +3474,10 @@ class Join(Op):
 
         rval = [grad_undefined(self, 0, axis)]
 
-        if 'float' in tensors[0].dtype or 'complex' in tensors[0].dtype:
+        dtypes = [as_tensor_variable(x).type.dtype for x in tensors]
+        out_dtype = scal.upcast(*dtypes)
+
+        if 'float' in out_dtype or 'complex' in out_dtype:
             # assume that this is differentiable
             split = Split(len(tensors))
             split_gz = split(gz, axis, stack(*[shape(x)[axis]
@@ -3491,6 +3494,7 @@ class Join(Op):
         else:
             # the output has integer type, so the gradient through it
             # is 0
+            # TODO what should be there?
             rval = rval + [tensor.zeros_like() for tensor in tensors]
 
         return rval

theano/tensor/tests/test_basic.py

@@ -3290,6 +3290,22 @@ class T_Join_and_Split(unittest.TestCase):
         utt.verify_grad(lambda a, b: join(1, a, b), [av, bv],
                         eps=1.0e-4, rel_tol=1.0e-3, mode=self.mode)
 
+    def test_join_matrix_dtypes(self):
+        # Test mixed dtype. There was a bug that caused crash in the past.
+        av = numpy.array([[1, 2, 3], [4, 5, 6]], dtype='int8')
+        bv = numpy.array([[7], [8]], dtype='float32')
+        a = self.shared(av)
+        b = as_tensor_variable(bv)
+        s = join(1, a, b)
+        want = numpy.array([[1, 2, 3, 7], [4, 5, 6, 8]], dtype='float32')
+        out = self.eval_outputs_and_check_join([s])
+        self.assertTrue((out == want).all())
+
+        grad(s.sum(), b)
+        grad(s.sum(), a)
+        utt.verify_grad(lambda b: join(1, a, b), [bv],
+                        eps=1.0e-4, rel_tol=1.0e-3, mode=self.mode)
+
     def test_join_matrix1_using_vertical_stack(self):
         a = self.shared(numpy.array([[1, 2, 3], [4, 5, 6]], dtype=self.floatX))
         b = as_tensor_variable(numpy.array([[7, 8, 9]], dtype=self.floatX))