Add legalization for TF.StopGradient to TF.Identity

PiperOrigin-RevId: 261734406

tensorflow/compiler/mlir/lite/tests/prepare-tf.mlir

@@ -287,3 +287,11 @@ func @matmulNoTransposeB(%arg0: tensor<1x1280xf32>, %arg1: tensor<1280x1000xf32>
   // CHECK: %7 = "tf.Transpose"(%arg1, %6) : (tensor<1280x1000xf32>, tensor<?xi32>) -> tensor<*xf32>
   // CHECK: %8 = "tf.MatMul"(%3, %7) {transpose_a = false, transpose_b = true} : (tensor<*xf32>, tensor<*xf32>) -> tensor<1x1000xf32>
 }
+
+func @stop_gradient(%arg0: tensor<3xi32>) -> tensor<3xi32> {
+  %0 = "tf.StopGradient"(%arg0) : (tensor<3xi32>) -> tensor<3xi32>
+  return %0 : tensor<3xi32>
+  // Should be converted to Identity and then from Identity to value
+  // CHECK-LABEL: stop_gradient
+  // CHECK:  return %arg0 : tensor<3xi32>
+}

tensorflow/compiler/mlir/lite/transforms/prepare_patterns.td

@@ -81,6 +81,8 @@ def : Pat<(TF_MatMulOp $a, $b, ConstBoolAttrTrue, $bt),
              /*delta=*/(ConstantOp TFi32<-1>)), (ConstantOp TFi32<1>))), $b,
            ConstBoolAttrFalse, $bt)>;
 
+def : Pat<(TF_StopGradientOp $arg), (TF_IdentityOp $arg)>;
+
 //===----------------------------------------------------------------------===//
 // Op removal patterns.
 //===----------------------------------------------------------------------===//

tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td

@@ -3138,6 +3138,42 @@ shape(squeeze(t, [2, 4])) ==> [1, 2, 3, 1]
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
 }
 
+def TF_StopGradientOp : TF_Op<"StopGradient", [NoSideEffect, SameOperandsAndResultType]> {
+  let summary = "Stops gradient computation.";
+
+  let description = [{
+When executed in a graph, this op outputs its input tensor as-is.
+
+When building ops to compute gradients, this op prevents the contribution of
+its inputs to be taken into account.  Normally, the gradient generator adds ops
+to a graph to compute the derivatives of a specified 'loss' by recursively
+finding out inputs that contributed to its computation.  If you insert this op
+in the graph it inputs are masked from the gradient generator.  They are not
+taken into account for computing gradients.
+
+This is useful any time you want to compute a value with TensorFlow but need
+to pretend that the value was a constant. Some examples include:
+
+*  The *EM* algorithm where the *M-step* should not involve backpropagation
+   through the output of the *E-step*.
+*  Contrastive divergence training of Boltzmann machines where, when
+   differentiating the energy function, the training must not backpropagate
+   through the graph that generated the samples from the model.
+*  Adversarial training, where no backprop should happen through the adversarial
+   example generation process.
+  }];
+
+  let arguments = (ins
+    TF_Tensor:$input
+  );
+
+  let results = (outs
+    TF_Tensor:$output
+  );
+
+  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
+}
+
 def TF_StridedSliceOp : TF_Op<"StridedSlice", [NoSideEffect]> {
   let summary = "Return a strided slice from `input`.";