int64_t out_row_dim = output_shape[output_shape.size() - 2];
  int64_t out_col_dim = output_shape[output_shape.size() - 1];

  int64_t expected_out_row_dim = op.getAdjX() ? x_col_dim : x_row_dim;
  int64_t expected_out_col_dim = op.getAdjY() ? y_row_dim : y_col_dim;

  if (expected_out_row_dim != ShapedType::kDynamic &&
      out_row_dim != ShapedType::kDynamic &&
      out_row_dim != expected_out_row_dim)

    Variadic<TF_Tensor>:$output
  );

  TF_DerivedOperandTypeListAttr Tin = TF_DerivedOperandTypeListAttr<1>;
  TF_DerivedResultTypeListAttr Tout = TF_DerivedResultTypeListAttr<0>;
  TF_DerivedResultShapeListAttr output_shapes = TF_DerivedResultShapeListAttr<0>;

  let hasCanonicalizer = 1;

  let hasVerifier = 1;


 let extraClassDeclaration = [{
    int num_branches() { return getBranches().size(); }

  return %arg0 : tensor<!tf_type.variant>
}

// CHECK-LABEL: iterators
func.func private @iterators(%arg0 : tensor<!tf_type.variant>) {
  %0 = "tf.Iterator"() {container = "", output_shapes = [#tf_type.shape<200x28x28x1>, #tf_type.shape<200x10>], output_types = [f32, f32], shared_name = "_iterator1"} : () -> tensor<!tf_type.resource>
  %1 = func.call @id(%arg0) : (tensor<!tf_type.variant>) -> tensor<!tf_type.variant>

    const xla::HloInputOutputAliasConfig& input_output_alias,
    absl::Span<const int> input_mapping,
    const std::map<int, const Tensor*>& resource_vars_snapshots,
    DataType output_dtype, const TensorShape& output_shape,
    Allocator* output_allocator, bool allocate_xla_tensors, se::Stream* stream,
    bool use_multiple_streams, std::shared_ptr<se::Event> definition_event) {

  return ConvertTensor(t, builder);
}

void ConvertToTensorShapeProto(ArrayRef<int64_t> shape,
                               TensorShapeProto* output_shape) {
  for (auto d : shape) {
    output_shape->add_dim()->set_size(ShapedType::isDynamic(d) ? kTFDynamicSize
                                                               : d);
  }
}

  %output_shape = arith.constant dense<[2, 3, 2, 2048]> : tensor<4xi32>
  %f16_weights = "tfl.pseudo_const"() {value = dense<1.0> : tensor<4x2x2x2048xf16>} : () -> tensor<4x2x2x2048xf16>
  %dq_weights = "tfl.dequantize"(%f16_weights) : (tensor<4x2x2x2048xf16>) -> tensor<4x2x2x2048xf32>
  %bias = "tfl.no_value"() {value} : () -> none

    %0 = "tf.TakeDataset"(%arg0, %cst) {device = "", metadata = "", output_shapes = [#tf_type.shape<>], output_types = [!tf_type.string]} : (tensor<!tf_type.variant>, tensor<i64>) -> tensor<!tf_type.variant>

      }
      auto output_shape =
          mlir::cast<RankedTensorType>(conv_op.getResult().getType())
              .getShape();
      SmallVector<int64_t, 4> transposed_output_shape = {
          output_shape[dnums.getOutputBatchDimension()],
          output_shape[dnums.getOutputSpatialDimensions().data()[0]],
          output_shape[dnums.getOutputSpatialDimensions().data()[1]],

      mlir::cast<ShapedType>(input.getType()).getShape();
  ArrayRef<int64_t> output_shape =
      mlir::cast<ShapedType>(output.getType()).getShape();

  int64_t agg_value = 1;
  for (size_t i = agg_start_idx; i < input_shape.size() - 1; ++i) {
    agg_value *= input_shape[i];
  }

  return (agg_value == output_shape[agg_start_idx]);
}

// Returns whether the given type `a` is broadcast-compatible with `b`.

      for (auto s : new_shape) {
        shape.push_back(
            builder.getI32IntegerAttr(mlir::TFL::ConvertToTfliteSize(s)));
      }
      auto output_shape = DenseElementsAttr::get(shape_type, shape);
      auto shape_op = builder.create<tfl::ConstOp>(loc, output_shape);
      op_state.addOperands({shape_op});
    }

    op_state.addTypes({type});
  }