remove sigmoid optimization for extreme values (#32)

tests/tensor/nnet/test_sigm.py

@@ -88,7 +88,8 @@ TestSoftplusBroadcast = makeBroadcastTester(
     ),
     good=dict(
         copymod(
-            _good_broadcast_unary_normal_no_complex, without=["uint8", "uint16"]
+            _good_broadcast_unary_normal_no_complex,
+            without=["uint8", "uint16", "big_scalar"],
         ),  # numpy function overflows with uint16.
         uint8=[
             np.arange(0, 89, dtype="uint8")

tests/tensor/test_basic.py

@@ -1290,6 +1290,7 @@ _good_broadcast_unary_normal_no_complex = dict(
     uint16=[np.arange(0, 89, dtype="uint16")],
     corner_case=[corner_case],
     empty=[np.asarray([], dtype=config.floatX)],
+    big_scalar=[np.arange(17.0, 29.0, 0.5, dtype=floatX)],
 )
 
 _grad_broadcast_unary_normal_no_complex = dict(

theano/tensor/nnet/sigm.py

@@ -80,15 +80,9 @@ class ScalarSigmoid(scalar.UnaryScalarOp):
             node.inputs[0].type == scalar.float32
             or node.inputs[0].type == scalar.float16
         ):
-            return (
-                """%(z)s = %(x)s < -88.0f ? 0.0 : %(x)s > 15.0f ? 1.0f : 1.0f /(1.0f + exp(-%(x)s));"""
-                % locals()
-            )
+            return """%(z)s = 1.0f / (1.0f + exp(-%(x)s));""" % locals()
         elif node.inputs[0].type == scalar.float64:
-            return (
-                """%(z)s = %(x)s < -709.0 ? 0.0 : %(x)s > 19.0 ? 1.0 : 1.0 /(1.0+exp(-%(x)s));"""
-                % locals()
-            )
+            return """%(z)s = 1.0 / (1.0 + exp(-%(x)s));""" % locals()
         else:
             raise NotImplementedError("only floatingpoint is implemented")