%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>

  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "data_format"
    value {
      s: "NHWC"
    }
  }
  attr {
    key: "dilations"
    value {
      list {
        i: 1
        i: 1
        i: 1
        i: 1
      }
    }
  }
  attr {
    key: "explicit_paddings"
    value {
      list {
      }

  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "data_format"
    value {
      s: "NCHW"
    }
  }
  attr {
    key: "dilations"
    value {
      list {
        i: 1
        i: 1
        i: 1
        i: 1
      }
    }
  }
  attr {
    key: "explicit_paddings"
    value {

        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)

  %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>

  %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>

  %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>

        mean, variance = scale, offset
        out = nn_ops.conv2d(
            input_tensor,
            self.filters,
            strides=strides,
            dilations=dilations,
            padding=padding,
            data_format='NHWC',
        )
        if has_bias:
          out = nn_ops.bias_add(out, self.bias, data_format='NHWC')
        if has_batch_norm:

      const int64_t dim = GetTensorSpatialDimIndex(num_dims, data_format, i);
      spatial_dim_indices.push_back(dim);

      const int64_t dilation = get_int(dilations[dim]);
      rhs_dilations.push_back(dilation);
      const int64_t stride = get_int(strides[dim]);
      window_strides.push_back(stride);

      Value pad_low, pad_high;