diff --git a/src/TensorFlowNET.Core/Gradients/gradients_impl.py.cs b/src/TensorFlowNET.Core/Gradients/gradients_impl.py.cs
index 83527399..fa9c9166 100644
--- a/src/TensorFlowNET.Core/Gradients/gradients_impl.py.cs
+++ b/src/TensorFlowNET.Core/Gradients/gradients_impl.py.cs
@@ -192,8 +192,8 @@ namespace Tensorflow
             {
                 if (!pending_count.ContainsKey(x.op.Name))
                     pending_count[x.op.Name] = 0;
-                else
-                    pending_count[x.op.Name] -= 1;
+
+                pending_count[x.op.Name] -= 1;
 
                 var ready = pending_count[x.op.Name] == 0;
 
@@ -288,7 +288,7 @@ namespace Tensorflow
                     }
                 }
                 if (is_stop_op)
-                    stop_ops.Add(op);
+                    stop_ops.Insert(0, op);
             }
             stop_ops.AddRange(stop_gradient_ops.Where(x => !stop_ops.Contains(x)));
             return stop_ops.ToArray();
diff --git a/src/TensorFlowNET.Core/Gradients/math_grad.py.cs b/src/TensorFlowNET.Core/Gradients/math_grad.py.cs
index 8377b72d..19d51463 100644
--- a/src/TensorFlowNET.Core/Gradients/math_grad.py.cs
+++ b/src/TensorFlowNET.Core/Gradients/math_grad.py.cs
@@ -1,5 +1,6 @@
 using System;
 using System.Collections.Generic;
+using System.Linq;
 using System.Text;
 
 namespace Tensorflow
@@ -13,8 +14,67 @@ namespace Tensorflow
         {
             var x = op.inputs[0];
             var y = op.inputs[1];
+            if (grad is Tensor && _ShapesFullySpecifiedAndEqual(x, y, grad))
+                return (grad, grad);
 
-            return (grad, grad);
+            var sx = array_ops.shape(x);
+            var sy = array_ops.shape(y);
+            var (rx, ry) = gen_array_ops.broadcast_gradient_args(sx, sy);
+
+            var r1 = gen_array_ops.reshape(math_ops.reduce_sum(grad, rx), sx);
+            var r2 = gen_array_ops.reshape(math_ops.reduce_sum(grad, ry), sy);
+
+            return (r1, r2);
+        }
+
+        public static (Tensor, Tensor) _IdGrad(Operation op, Tensor grad)
+        {
+            return (grad, null);
+        }
+
+        public static (Tensor, Tensor) _MulGrad(Operation op, Tensor grad)
+        {
+            var x = op.inputs[0];
+            var y = op.inputs[1];
+            if (grad is Tensor && _ShapesFullySpecifiedAndEqual(x, y, grad) &&
+                new TF_DataType[] { tf.int32, tf.float32 }.Contains(grad.dtype))
+                return (gen_math_ops.mul(grad, y), gen_math_ops.mul(grad, x));
+
+            var sx = array_ops.shape(x);
+            var sy = array_ops.shape(y);
+            var (rx, ry) = gen_array_ops.broadcast_gradient_args(sx, sy);
+
+            x = math_ops.conj(x);
+            y = math_ops.conj(y);
+
+            var r1 = math_ops.reduce_sum(gen_math_ops.mul(grad, y), rx);
+            var r2 = math_ops.reduce_sum(gen_math_ops.mul(x, grad), ry);
+
+            return (gen_array_ops.reshape(r1, sx), gen_array_ops.reshape(r2, sy));
+        }
+
+        public static (Tensor, Tensor) _SubGrad(Operation op, Tensor grad)
+        {
+            var x = op.inputs[0];
+            var y = op.inputs[1];
+            if (grad is Tensor && _ShapesFullySpecifiedAndEqual(x, y, grad))
+                return (grad, -grad);
+
+            var sx = array_ops.shape(x);
+            var sy = array_ops.shape(y);
+            var (rx, ry) = gen_array_ops.broadcast_gradient_args(sx, sy);
+
+            var r1 = gen_array_ops.reshape(math_ops.reduce_sum(grad, rx), sx);
+            var r2 = gen_array_ops.reshape(-math_ops.reduce_sum(grad, ry), sy);
+
+            return (r1, r2);
+        }
+
+        public static bool _ShapesFullySpecifiedAndEqual(Tensor x, Tensor y, Tensor grad)
+        {
+            return false;
+            /*return string.Join(",", x.shape).Equals(string.Join(",", y.shape)) &&
+                   string.Join(",", x.shape).Equals(string.Join(",", grad.shape));*/
         }
 
         public static (Tensor, Tensor) _SumGrad(Operation op, Tensor grad)
@@ -71,8 +131,7 @@ namespace Tensorflow
             x = math_ops.conj(x);
             y = math_ops.conj(y);
             y = math_ops.conj(z);
-
-            var gx = gen_array_ops.reshape(math_ops.reduce_sum(grad * y * gen_math_ops.pow(x, y - 1), rx), sx);
+            var gx = gen_array_ops.reshape(math_ops.reduce_sum(grad * y * gen_math_ops.pow(x, y - 1.0), rx), sx);
             Tensor log_x = null;
             // Avoid false singularity at x = 0
             if (x.dtype.is_complex())
@@ -81,7 +140,9 @@ namespace Tensorflow
             }
             else
             {
-                log_x = array_ops.where(x > 0, gen_array_ops.log(x), array_ops.zeros_like(x));
+                var x1 = gen_array_ops.log(x);
+                var y1 = array_ops.zeros_like(x);
+                log_x = array_ops.where(x > 0.0, x1, y1);
             }
 
             var gy = gen_array_ops.reshape(math_ops.reduce_sum(grad * z * log_x, ry), sy);
diff --git a/src/TensorFlowNET.Core/Operations/array_ops.py.cs b/src/TensorFlowNET.Core/Operations/array_ops.py.cs
index 62722a32..23a6eef3 100644
--- a/src/TensorFlowNET.Core/Operations/array_ops.py.cs
+++ b/src/TensorFlowNET.Core/Operations/array_ops.py.cs
@@ -147,7 +147,27 @@ namespace Tensorflow
 
         public static Tensor zeros_like(Tensor tensor, TF_DataType dtype = TF_DataType.DtInvalid, string name = "", bool optimize = true)
         {
-            throw new NotImplementedException("zeros_like");
+            return Python.with<ops.name_scope, Tensor>(new ops.name_scope(name, "zeros_like", new Tensor[] { tensor }), scope =>
+            {
+                name = scope;
+                tensor = ops.convert_to_tensor(tensor, name: "tensor");
+
+                // is_fully_defined return unexpected value.
+                if (optimize && tensor_util.to_shape(tensor.shape).is_fully_defined() && dtype != TF_DataType.TF_VARIANT)
+                {
+
+                }
+
+                if(dtype != TF_DataType.DtInvalid && dtype != tensor.dtype && dtype != TF_DataType.TF_VARIANT)
+                {
+                    throw new NotImplementedException("zeros_like");
+                    // return zeros(shape_internal(tensor, optimize: optimize), dtype: dtype, name: name);
+                }
+                else
+                {
+                    return gen_array_ops.zeros_like(tensor, name: name);
+                }
+            });
         }
     }
 }
diff --git a/src/TensorFlowNET.Core/Operations/gen_array_ops.cs b/src/TensorFlowNET.Core/Operations/gen_array_ops.cs
index 7093967b..efc8e4e0 100644
--- a/src/TensorFlowNET.Core/Operations/gen_array_ops.cs
+++ b/src/TensorFlowNET.Core/Operations/gen_array_ops.cs
@@ -106,9 +106,9 @@ namespace Tensorflow
             throw new NotImplementedException("where");
         }
 
-        public static Tensor select(Tensor condition, Tensor x, Tensor y, string name = "")
+        public static Tensor select(Tensor condition, Tensor t, Tensor e, string name = "")
         {
-            var _op = _op_def_lib._apply_op_helper("Select", name, new { condition, x, y });
+            var _op = _op_def_lib._apply_op_helper("Select", name, new { condition, t, e });
             return _op.outputs[0];
         }
 
diff --git a/src/TensorFlowNET.Core/Operations/gen_math_ops.cs b/src/TensorFlowNET.Core/Operations/gen_math_ops.cs
index 627f9850..57921209 100644
--- a/src/TensorFlowNET.Core/Operations/gen_math_ops.cs
+++ b/src/TensorFlowNET.Core/Operations/gen_math_ops.cs
@@ -24,7 +24,7 @@ namespace Tensorflow
             return _op.outputs[0];
         }
 
-        public static Tensor sub(Tensor x, Tensor y, string name = "")
+        public static Tensor sub<Tx, Ty>(Tx x, Ty y, string name = "")
         {
             var _op = _op_def_lib._apply_op_helper("Sub", name, args: new { x, y });
 
@@ -89,7 +89,7 @@ namespace Tensorflow
             return _op.outputs[0];
         }
 
-        public static Tensor pow(Tensor x, double y, string name = "")
+        public static Tensor pow<Tx, Ty>(Tx x, Ty y, string name = "")
         {
             var _op = _op_def_lib._apply_op_helper("Pow", name, args: new { x, y });
 
diff --git a/src/TensorFlowNET.Core/Tensors/Tensor.Operators.cs b/src/TensorFlowNET.Core/Tensors/Tensor.Operators.cs
index a2851de8..77c5c5c0 100644
--- a/src/TensorFlowNET.Core/Tensors/Tensor.Operators.cs
+++ b/src/TensorFlowNET.Core/Tensors/Tensor.Operators.cs
@@ -23,15 +23,10 @@ namespace Tensorflow
             });
         }
 
-        public static Tensor operator -(Tensor t1)
-        {
-            return gen_math_ops.neg(t1);
-        }
-
-        public static Tensor operator -(Tensor t1, Tensor t2)
-        {
-            return gen_math_ops.sub(t1, t2);
-        }
+        public static Tensor operator -(Tensor t1) => gen_math_ops.neg(t1);
+        public static Tensor operator -(Tensor t1, Tensor t2) => gen_math_ops.sub(t1, t2);
+        public static Tensor operator -(Tensor t1, int t2) => gen_math_ops.sub(t1, t2);
+        public static Tensor operator -(Tensor t1, double t2) => gen_math_ops.sub(t1, t2);
 
         public static Tensor operator *(double x, Tensor y)
         {
@@ -85,14 +80,9 @@ namespace Tensorflow
             });
         }
 
-        public static Tensor operator >(Tensor x, int y)
-        {
-            return gen_array_ops.greater(x, y);
-        }
-
-        public static Tensor operator <(Tensor x, int y)
-        {
-            return gen_array_ops.less(x, y);
-        }
+        public static Tensor operator >(Tensor x, int y) => gen_array_ops.greater(x, y);
+        public static Tensor operator >(Tensor x, double y) => gen_array_ops.greater(x, y);
+        public static Tensor operator <(Tensor x, int y) => gen_array_ops.less(x, y);
+        public static Tensor operator <(Tensor x, double y) => gen_array_ops.less(x, y);
     }
 }
diff --git a/src/TensorFlowNET.Core/ops.py.cs b/src/TensorFlowNET.Core/ops.py.cs
index 52600d29..d98854f6 100644
--- a/src/TensorFlowNET.Core/ops.py.cs
+++ b/src/TensorFlowNET.Core/ops.py.cs
@@ -294,8 +294,16 @@ namespace Tensorflow
                 {
                     case "Add":
                         return math_grad._AddGrad(oper, out_grads);
+                    case "Identity":
+                        return math_grad._IdGrad(oper, out_grads);
+                    case "Mul":
+                        return math_grad._MulGrad(oper, out_grads);
                     case "Sum":
                         return math_grad._SumGrad(oper, out_grads);
+                    case "Sub":
+                        return math_grad._SubGrad(oper, out_grads);
+                    case "Pow":
+                        return math_grad._PowGrad(oper, out_grads);
                     case "RealDiv":
                         return math_grad._RealDivGrad(oper, out_grads);
                     default: