diff --git a/src/TensorFlowNET.Core/Gradients/math_grad.py.cs b/src/TensorFlowNET.Core/Gradients/math_grad.cs
similarity index 87%
rename from src/TensorFlowNET.Core/Gradients/math_grad.py.cs
rename to src/TensorFlowNET.Core/Gradients/math_grad.cs
index 5d02033c..4380a7fa 100644
--- a/src/TensorFlowNET.Core/Gradients/math_grad.py.cs
+++ b/src/TensorFlowNET.Core/Gradients/math_grad.cs
@@ -8,7 +8,7 @@ namespace Tensorflow.Gradients
     /// <summary>
     /// Gradients for operators defined in math_ops.py.
     /// </summary>
-    public class math_grad
+    public class math_grad : Python
     {
         public static Tensor[] _AddGrad(Operation op, Tensor[] grads)
         {
@@ -33,6 +33,16 @@ namespace Tensorflow.Gradients
             return new Tensor[] { grads[0] };
         }
 
+        public static Tensor[] _LogGrad(Operation op, Tensor[] grads)
+        {
+            var grad = grads[0];
+            var x = op.inputs[0];
+            return with(ops.control_dependencies(new Operation[] { grad }), dp => {
+                x = math_ops.conj(x);
+                return new Tensor[] { grad * math_ops.reciprocal(x) };
+            });
+        }
+
         public static Tensor[] _MulGrad(Operation op, Tensor[] grads)
         {
             var x = op.inputs[0];
@@ -106,6 +116,11 @@ namespace Tensorflow.Gradients
             return new Tensor[] { math_ops.truediv(sum_grad, math_ops.cast(factor, sum_grad.dtype)), null };
         }
 
+        public static Tensor[] _NegGrad(Operation op, Tensor[] grads)
+        {
+            return new Tensor[] { -grads[0] };
+        }
+
         private static Tensor _safe_shape_div(Tensor x, Tensor y)
         {
             return math_ops.floordiv(x, gen_math_ops.maximum(y, 1));
@@ -145,13 +160,16 @@ namespace Tensorflow.Gradients
                 var axes = tensor_util.constant_value(op.inputs[1]);
                 if(!(axes is null))
                 {
-                    var rank = axes.shape.Rank;
-                    grad = array_ops.reshape(grad, new int[] { 1 });
-                    if (!input_0_shape.Contains(-1))
-                        input_shape = constant_op.constant(input_0_shape);
-                    else
-                        input_shape = array_ops.shape(op.inputs[0]);
-                    return new Tensor[] { gen_array_ops.tile(grad, input_shape), null };
+                    var rank = input_0_shape.Length;
+                    if (Enumerable.SequenceEqual(Enumerable.Range(0, rank), axes.Data<int>()))
+                    {
+                        grad = array_ops.reshape(grad, new int[] { 1 });
+                        if (!input_0_shape.Contains(-1))
+                            input_shape = constant_op.constant(input_0_shape);
+                        else
+                            input_shape = array_ops.shape(op.inputs[0]);
+                        return new Tensor[] { gen_array_ops.tile(grad, input_shape), null };
+                    }
                 }
             }
 
diff --git a/src/TensorFlowNET.Core/Gradients/nn_grad.py.cs b/src/TensorFlowNET.Core/Gradients/nn_grad.py.cs
index a4840fd7..6740bdbf 100644
--- a/src/TensorFlowNET.Core/Gradients/nn_grad.py.cs
+++ b/src/TensorFlowNET.Core/Gradients/nn_grad.py.cs
@@ -27,6 +27,21 @@ namespace Tensorflow.Gradients
             return new Tensor[] { gen_nn_ops.relu_grad(grads[0], op.outputs[0]) };
         }
 
+        /// <summary>
+        /// The derivative of the softmax nonlinearity.
+        /// </summary>
+        /// <param name="op"></param>
+        /// <param name="grads"></param>
+        /// <returns></returns>
+        public static Tensor[] _SoftmaxGrad(Operation op, Tensor[] grads)
+        {
+            var grad_softmax = grads[0];
+
+            var softmax = op.outputs[0];
+            var sum_channels = math_ops.reduce_sum(grad_softmax * softmax, -1, keepdims: true);
+            return new Tensor[] { (grad_softmax - sum_channels) * softmax };
+        }
+
         /// <summary>
         /// Gradient function for SoftmaxCrossEntropyWithLogits.
         /// </summary>
diff --git a/src/TensorFlowNET.Core/Gradients/ops.gradient_function_mapping.cs b/src/TensorFlowNET.Core/Gradients/ops.gradient_function_mapping.cs
index 9601077b..7b650d00 100644
--- a/src/TensorFlowNET.Core/Gradients/ops.gradient_function_mapping.cs
+++ b/src/TensorFlowNET.Core/Gradients/ops.gradient_function_mapping.cs
@@ -24,6 +24,8 @@ namespace Tensorflow
                         return nn_grad._BiasAddGrad(oper, out_grads);
                     case "Identity":
                         return math_grad._IdGrad(oper, out_grads);
+                    case "Log":
+                        return math_grad._LogGrad(oper, out_grads);
                     case "MatMul":
                         return math_grad._MatMulGrad(oper, out_grads);
                     case "Merge":
@@ -32,6 +34,8 @@ namespace Tensorflow
                         return math_grad._MulGrad(oper, out_grads);
                     case "Mean":
                         return math_grad._MeanGrad(oper, out_grads);
+                    case "Neg":
+                        return math_grad._NegGrad(oper, out_grads);
                     case "Sum":
                         return math_grad._SumGrad(oper, out_grads);
                     case "Sub":
@@ -46,6 +50,8 @@ namespace Tensorflow
                         return nn_grad._ReluGrad(oper, out_grads);
                     case "Squeeze":
                         return array_grad._SqueezeGrad(oper, out_grads);
+                    case "Softmax":
+                        return nn_grad._SoftmaxGrad(oper, out_grads);
                     case "SoftmaxCrossEntropyWithLogits":
                         return nn_grad._SoftmaxCrossEntropyWithLogitsGrad(oper, out_grads);
                     case "Transpose":
diff --git a/src/TensorFlowNET.Core/Operations/gen_math_ops.cs b/src/TensorFlowNET.Core/Operations/gen_math_ops.cs
index 1ccb0e96..4cfdc7cf 100644
--- a/src/TensorFlowNET.Core/Operations/gen_math_ops.cs
+++ b/src/TensorFlowNET.Core/Operations/gen_math_ops.cs
@@ -142,6 +142,13 @@ namespace Tensorflow
             return _op.outputs[0];
         }
 
+        public static Tensor reciprocal(Tensor x, string name = null)
+        {
+            var _op = _op_def_lib._apply_op_helper("Reciprocal", name, args: new { x });
+
+            return _op.outputs[0];
+        }
+
         public static Tensor floor_mod(Tensor x, Tensor y, string name = null)
         {
             var _op = _op_def_lib._apply_op_helper("FloorMod", name, args: new { x, y });
diff --git a/src/TensorFlowNET.Core/Operations/math_ops.py.cs b/src/TensorFlowNET.Core/Operations/math_ops.py.cs
index 3c6ea594..c8e3f98f 100644
--- a/src/TensorFlowNET.Core/Operations/math_ops.py.cs
+++ b/src/TensorFlowNET.Core/Operations/math_ops.py.cs
@@ -126,6 +126,16 @@ namespace Tensorflow
             return gen_data_flow_ops.dynamic_stitch(a1, a2);
         }
 
+        /// <summary>
+        /// Computes the reciprocal of x element-wise.
+        /// </summary>
+        /// <param name="x"></param>
+        /// <param name="name"></param>
+        /// <returns></returns>
+        public static Tensor reciprocal(Tensor x, string name = null)
+            => gen_math_ops.reciprocal(x, name: name);
+        
+
         /// <summary>
         /// Computes log(sum(exp(elements across dimensions of a tensor))).
         /// Reduces `input_tensor` along the dimensions given in `axis`.