%cst_0 = arith.constant dense<1> : tensor<3xi32>
  %0 = "tf.Squeeze"(%arg0) {T = f32, _output_shapes = ["tfshape$dim { size: 4 } dim { size: 64 } dim { size: 64 }"], device = "", squeeze_dims = []} : (tensor<4x64x64x1xf32>) -> tensor<4x64x64xf32>

   def FuseBinaryOpWithTransposeConv#binaryOp : Pat<
    (binaryOp (TFL_TransposeConvOp:$output $output_shape, $weights, $input,
                (Arith_ConstantOp FloatElementsAttr:$bias), $padding,
                $stride_h, $stride_w, TFL_AF_None),
              (Arith_ConstantOp FloatElementsAttr:$value), $act_fn),
    (TFL_TransposeConvOp $output_shape, $weights, $input,
      (binaryOp (Arith_ConstantOp $bias),

    auto new_shape = rewriter.create<TF::ShapeOp>(loc, shape_type, input);
    SmallVector<int64_t, 8> output_shape(/*Size=*/1, op.getNumElements());
    for (const auto &dim : dense_elem_attr.getValues<APInt>())
      output_shape.push_back(dim.getSExtValue());
    RankedTensorType result_type = tensorflow::GetTypeFromTFTensorShape(
        output_shape, getElementTypeOrSelf(input));
    rewriter.replaceOpWithNewOp<TF::ReshapeOp>(op, result_type, input,

      auto begin =
          rewriter.create<ConstOp>(op->getLoc(), axis_type, begin_attr);

      SmallVector<int64_t, 4> output_shape;
      output_shape.append(input_shape.begin(), input_shape.end());
      output_shape[axis_i] = size_i;
      auto size_attr = DenseIntElementsAttr::get(axis_type, output_shape);
      auto size = rewriter.create<ConstOp>(op->getLoc(), axis_type, size_attr);

    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(); }

  // n_input_entries/n_output_entries
  // = group_size
  auto input_shape = Shape(scope, op.input(0));
  auto output_shape = Shape(scope, op.output(0));
  auto zero = Const(scope, 0);
  auto group_size = SafeDivHelper(scope, Prod(scope, input_shape, zero),
                                  Prod(scope, output_shape, zero));

  // propagate sum_grad/group_size
  grad_outputs->push_back(

      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});
  }

  // CHECK-DAG: [[OUTPUT_SHAPE:%.+]] = "tf.ConcatV2"([[OUTPUT_BATCH]], [[OUTER_SHAPE_0]], [[OUTER_SHAPE_1]], [[PADDED_SHAPE_SPLITS]]#3, [[ZERO_I64]])
  // CHECK-DAG: [[RESHAPED:%.+]] = "tf.Reshape"([[PADDED]], [[RESHAPED_SHAPE]])
  // CHECK-DAG: [[PERMUTED:%.+]] = "tf.Transpose"([[RESHAPED]], [[PERMUTATION]])
  // CHECK-DAG: [[RESULT:%.+]] = "tf.Reshape"([[PERMUTED]], [[OUTPUT_SHAPE]])
  // CHECK-DAG: return [[RESULT]]

    %arg1: tensor<i64>
  ) {
    %1 = "tf.ReduceDataset"(%arg0, %arg1) {
      Targuments = [],
      Tstate = [i64], device = "",
      f = @__reduce_func_1, f._tf_data_function = true,
      output_shapes = [#tf_type.shape<>],
      output_types = [i64], use_inter_op_parallelism = true, _xla_compile_device_type="TPU"} :
 (tensor<!tf_type.variant>, tensor<i64>) -> (tensor<i64>)
    func.return
 }
 ```

[i32, i32, i32, !tf_type.variant, !tf_type.variant, f32], _lower_using_switch_merge = true, _num_original_outputs = 6 : i64, _read_only_resource_inputs = [], body = @map_while_body_180, cond = @map_while_cond_170, device = "", is_stateless = true, output_shapes = [#tf_type.shape<>, #tf_type.shape<>, #tf_type.shape<>, #tf_type.shape<>, #tf_type.shape<>, #tf_type.shape<?>], parallel_iterations = 4 : i64, shape_invariant} : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<!tf_type.variant<tensor<*xf32>>>,...