Fix NaN value comparisons in relu, max and min ops (apache#14262)

* nan comparison

* fix relu grad

src/operator/mshadow_op.h

@@ -127,10 +127,6 @@ MXNET_UNARY_MATH_OP(softsign, a / (1.0f + math::fabs(a)));
 
 MXNET_UNARY_MATH_OP(softsign_grad, 1.0f /  math::sqr(1.0f + math::fabs(a)));
 
-MXNET_UNARY_MATH_OP_NC(relu, a > DType(0) ? a : DType(0));
-
-MXNET_UNARY_MATH_OP_NC(relu_grad, a > DType(0) ? DType(1) : DType(0));
-
 MXNET_UNARY_MATH_OP_NC(selu, DType(SELU_LAMBDA) *
                          (a > DType(0) ? a : DType(math::id(SELU_ALPHA) * math::expm1(a))));
 
@@ -317,12 +313,6 @@ MXNET_BINARY_MATH_OP(rpower, math::pow(b, a));
 
 MXNET_BINARY_MATH_OP(rpower_grad, math::id(a) * math::log(b));
 
-/*! \brief used for generate element of maximum */
-MXNET_BINARY_MATH_OP(maximum, a > b ? a : b);
-
-/*! \brief used for generate element of minimum */
-MXNET_BINARY_MATH_OP_NC(minimum, a < b ? a : b);
-
 MXNET_UNARY_MATH_OP_NC(nt, a != DType(0) ? DType(0) : DType(1));
 
 MXNET_BINARY_MATH_OP_NC(ge, a >= b ? DType(1) : DType(0));
@@ -788,6 +778,44 @@ namespace isnan_typed {
   }
 };  // namespace isnan_typed
 
+MXNET_UNARY_MATH_OP_NC(relu, isnan_typed::IsNan(a) || (a > DType(0)) ? a : DType(0));
+
+/*! \brief used for computing gradient of relu operator */
+struct relu_grad : public mxnet_op::tunable {
+  template<typename DType>
+  MSHADOW_XINLINE static DType Map(DType a) {
+    if (isnan_typed::IsNan(a)) {
+      return a;
+    } else {
+      return a > DType(0) ? DType(1) : DType(0);
+    }
+  }
+};
+
+/*! \brief used for computing binary operator maximum */
+struct maximum : public mxnet_op::tunable {
+  template<typename DType>
+  MSHADOW_XINLINE static DType Map(DType a, DType b) {
+    if (isnan_typed::IsNan(a)) {
+      return a;
+    } else {
+      return (a > b ? a : b);
+    }
+  }
+};
+
+/*! \brief used for computing binary operator minimum */
+struct minimum : public mxnet_op::tunable {
+  template<typename DType>
+  MSHADOW_XINLINE static DType Map(DType a, DType b) {
+    if (isnan_typed::IsNan(a)) {
+      return a;
+    } else {
+      return DType(a < b ? a : b);
+    }
+  }
+};
+
 /*! \brief sum reducer that ignores NaN values in the input */
 struct nansum {
   /*! \brief do reduction into dst */

tests/python/unittest/test_ndarray.py

@@ -1550,6 +1550,35 @@ def test_ndarray_is_nan():
     np.testing.assert_equal(output.asnumpy(), expected_output.astype(int))
     # astype since numpy functions default return type is boolean array instead of int
 
+@with_seed()
+def test_ndarray_nan_comparison():
+    random_dimensions = np.random.randint(2, 5)
+    random_shape = [np.random.randint(2, 5) for i in range(random_dimensions)]
+    data1 = mxnet.test_utils.rand_ndarray(random_shape,'default')
+    data2 = mxnet.test_utils.rand_ndarray(random_shape,'default')
+    data1[1][0] = np.NaN
+    data2[0][0] = np.NaN
+
+    nd_max = mx.nd.maximum(data1, data2)
+    np_max = np.maximum(data1.asnumpy(), data2.asnumpy())
+    np.testing.assert_equal(nd_max.asnumpy(), np_max)
+
+    nd_min = mx.nd.minimum(data1, data2)
+    np_min = np.minimum(data1.asnumpy(), data2.asnumpy())
+    np.testing.assert_equal(nd_min.asnumpy(), np_min)
+
+    nd_relu = mx.nd.relu(data1)
+    np_relu = np.maximum(data1.asnumpy(), 0)
+    np.testing.assert_equal(nd_relu.asnumpy(), np_relu)
+
+    data1.attach_grad()
+    with mx.autograd.record():
+        y = mx.nd.relu(data1)
+    y.backward()
+    data1_grad = data1.grad.asnumpy()
+    for i in (np.isnan(data1_grad))[1][0].flatten():
+        assert i == True
+
 if __name__ == '__main__':
     import nose
     nose.runmodule()