%0 = "tf.Conv2D"(%arg0, %arg1) {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x2x2x3xf32>, tensor<2x2x3x2xf32>) -> tensor<*xf32>

  %rst = "tf.Conv2D"(%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<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

                                     Value input, Value filter,
                                     int8_t input_zp_value, ArrayAttr strides,
                                     ArrayAttr dilations,
                                     StringAttr conv_padding,
                                     ArrayAttr explicit_paddings,
                                     Value &padding, int num_dims = 4);

  %6 = "tf.Cast"(%5) {Truncate = false} : (tensor<1x100x100x1xi8>) -> tensor<1x100x100x1xf32>

  %7 = "tf.Sub"(%6, %cst_4) : (tensor<1x100x100x1xf32>, tensor<f32>) -> tensor<1x100x100x1xf32>
  %8 = "tf.Conv2D"(%7, %cst_3) {dilations = [1, 1, 1, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x100x100x1xf32>, tensor<3x3x1x1xf32>) -> tensor<1x98x98x1xf32>
  %9 = "tf.AddV2"(%8, %cst_6) : (tensor<1x98x98x1xf32>, tensor<1xf32>) -> tensor<1x98x98x1xf32>

    %2 = "tf.Conv2D"(%1, %0) {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

        offset = [0.5] * self.out_channel_size
        mean, variance = scale, offset
        out = nn_ops.conv2d(
            input_tensor,
            self.filters,
            strides=strides,
            dilations=dilations,
            padding=padding,
            data_format='NHWC',
            name='sample/conv',
        )
        if bias_fn is not None:
          out = nn_ops.bias_add(out, self.bias)

  %4 = "tf.Transpose"(%3, %cst_1) {_tpu_replicate = "cluster", device = ""} : (tensor<1x3x3x4xbf16>, tensor<4xi32>) -> tensor<1x3x4x3xbf16>
  %5 = "tf.Conv2D"(%4, %cst) {_tpu_replicate = "cluster", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xbf16>, tensor<2x3x3x2xbf16>) -> tensor<1x3x2x2xbf16>

  %2 = "tf.Conv2D"(%1, %0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

  %0 = "tf.Conv2D"(%output, %cst) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true}> {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", device = ""} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>

        %w = "tf.Mul"(%cst, %cst_2) : (tensor<2x3x3x1xf32>, tensor<f32>) -> tensor<2x3x3x1xf32>
        %0 = "tf.DepthwiseConv2dNative"(%arg0, %w) {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>