"""
        scale = [1.0] * self.out_channel_size
        offset = [0.5] * self.out_channel_size
        mean, variance = scale, offset
        out = nn_ops.depthwise_conv2d_native(
            input_tensor,
            self.filters,
            strides=[1, 2, 2, 1],
            dilations=[1, 1, 1, 1],
            padding='SAME',
            data_format='NHWC',
        )

    func.return %1: tensor<*xf32>
  }
  func.func private @composite_depthwise_conv2d_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {
      attr_map = "0:strides,1:padding,2:explicit_paddings,3:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1]

    func.return %1: tensor<*xf32>
  }
  func.func private @composite_depthwise_conv2d_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {
      attr_map = "0:strides,1:padding,2:explicit_paddings,3:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1]

  %fq = "tf.FakeQuantWithMinMaxVars"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<f32>, tensor<f32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.DepthwiseConv2dNative"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x30x30x16xf32>

  %cst = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
  %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
  %0 = "tf.DepthwiseConv2dNative"(%arg0, %cst) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<*xf32>, tensor<2x3x3x1xf32>) -> tensor<?x?x?x3xf32>

// CHECK-NEXT: %[[DEPTHWISECONV2D_0:.*]] = "tf.DepthwiseConv2dNative"(%arg0, %arg1)
// CHECK-SAME: attr_map = "0:strides,1:padding,2:explicit_paddings,3:dilations"
// CHECK-NEXT: return %[[DEPTHWISECONV2D_0:.*]]

// CHECK-LABEL: private @composite_depthwise_conv2d_fn_1
// CHECK-NEXT: %[[DEPTHWISECONV2D_0:.*]] = "tf.DepthwiseConv2dNative"(%arg0, %arg1)

  %fq = "tf.FakeQuantWithMinMaxVars"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<f32>, tensor<f32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.DepthwiseConv2dNative"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x30x30x16xf32>

  %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1]
  } : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<*xf32>
  %1 = "tf.Relu6"(%0) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>


  %3 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {

    func.return %1: tensor<*xf32>
  }
  func.func private @composite_depthwise_conv2d_fn(%arg0: tensor<1x3x4x512xf32>, %arg1: tensor<2x3x3x512xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {
      attr_map = "0:strides,1:padding,2:explicit_paddings,3:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1]

  %cst_1 = arith.constant dense<0> : tensor<2x2xi32>
  %0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
  %1 = "tf.DepthwiseConv2dNative"(%0, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>