Merge pull request #5289 from nouiz/isnan

theano/scalar/basic.py

@@ -1334,13 +1334,17 @@ class IsNan(FixedLogicalComparison):
         (z,) = outputs
         if node.inputs[0].type in complex_types:
             raise NotImplementedError()
+        # Discrete type can never be nan
+        if node.inputs[0].type in discrete_types:
+            return "%(z)s = false;" % locals()
+
         # Windows tries to be different and sometimes return -1, but we want
         # to be consistent with numpy (which returns True), hence the "abs".
         return "%(z)s = abs(isnan(%(x)s));" % locals()
 
     def c_code_cache_version(self):
         scalarop_version = super(IsNan, self).c_code_cache_version()
-        return tuple(scalarop_version) + (2,)
+        return tuple(scalarop_version) + (3,)
 isnan = IsNan()
 
 
@@ -1355,10 +1359,18 @@ class IsInf(FixedLogicalComparison):
         (z,) = outputs
         if node.inputs[0].type in complex_types:
             raise NotImplementedError()
+        # Discrete type can never be inf
+        if node.inputs[0].type in discrete_types:
+            return "%(z)s = false;" % locals()
+
         # Note that the C isinf returns -1 for -Inf and +1 for +Inf, while
         # numpy simply returns True: we mimic numpy's behavior here, thus
         # the absolute value.
         return "%(z)s = abs(isinf(%(x)s));" % locals()
+
+    def c_code_cache_version(self):
+        scalarop_version = super(IsInf, self).c_code_cache_version()
+        return tuple(scalarop_version) + (3,)
 isinf = IsInf()
 
 

theano/tensor/basic.py

@@ -1825,12 +1825,39 @@ def neq(a, b):
 def isnan(a):
     """isnan(a)"""
 
+# Rename isnan to isnan_ to allow to bypass it when not needed.
+# glibc 2.23 don't allow isnan on int, so we remove it from the graph.
+isnan_ = isnan
+
+
+def isnan(a):
+    """isnan(a)"""
+    a = as_tensor_variable(a)
+    if a.dtype in discrete_dtypes:
+        return alloc(numpy.asarray(False, dtype="bool"),
+                     *[a.shape[i] for i in range(a.ndim)])
+    return isnan_(a)
+
 
 @_scal_elemwise
 def isinf(a):
     """isinf(a)"""
 
 
+# Rename isnan to isnan_ to allow to bypass it when not needed.
+# glibc 2.23 don't allow isnan on int, so we remove it from the graph.
+isinf_ = isinf
+
+
+def isinf(a):
+    """isinf(a)"""
+    a = as_tensor_variable(a)
+    if a.dtype in discrete_dtypes:
+        return alloc(numpy.asarray(False, dtype="bool"),
+                     *[a.shape[i] for i in range(a.ndim)])
+    return isinf_(a)
+
+
 def allclose(a, b, rtol=1.e-5, atol=1.e-8, equal_nan=False):
     """
     Implement Numpy's ``allclose`` on tensors.

theano/tensor/tests/test_basic.py

@@ -2860,6 +2860,21 @@ def test_nan_inf_constant_signature():
     assert f(numpy.nan) == 0
 
 
+def test_isnan():
+    for x in [tensor.matrix(), tensor.imatrix(), tensor.matrix(dtype='bool')]:
+        y = tensor.isnan(x)
+        assert isinstance(y.owner.op, tensor.Elemwise) == (
+            x.dtype not in tensor.discrete_dtypes)
+        assert y.dtype == 'bool'
+
+        # Test c code generator even for int type.
+        y = tensor.isnan_(x)
+        assert isinstance(y.owner.op, tensor.Elemwise)
+        assert y.dtype == 'bool'
+        f = theano.function([x], y, allow_input_downcast=True)
+        f([[0, 1, 2]])
+
+
 class T_Shape(unittest.TestCase):
     def test_basic0(self):
         s = shape(numpy.ones((5, 3)))