Conv2D Operator Definition

The following uses a complex Conv2D as an example to describe how to define an operator.

Conv2D operator prototype definition:

REG_OP(Conv2D)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT, DT_DOUBLE, DT_INT8}))
    .INPUT(filter, TensorType({DT_FLOAT16, DT_FLOAT, DT_DOUBLE, DT_INT8}))
    .OPTIONAL_INPUT(bias, TensorType({DT_FLOAT16, DT_FLOAT, DT_DOUBLE, DT_INT32}))
    .OPTIONAL_INPUT(offset_w, TensorType({DT_INT8}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT, DT_DOUBLE, DT_INT32}))
    .REQUIRED_ATTR(strides, ListInt)
    .REQUIRED_ATTR(pads, ListInt)
    .ATTR(dilations, ListInt, {1, 1, 1, 1})
    .ATTR(groups, Int, 1)
    .ATTR(data_format, String, "NHWC")
    .ATTR(offset_x, Int, 0)
    .OP_END_FACTORY_REG(Conv2D)

The prototype definition shows that the Conv2D operator has two required inputs (x and filter), two optional inputs (bias and offset_w), two required attributes (strides and pads), and four optional attributes (dilations, groups, data_format, and offset_x). The definition code of Conv2D is as follows.

auto conv2d = op::Conv2D("Conv2d")
    .set_input_x(quant)
    .set_input_filter(conv_weight)
    .set_input_bias(conv_bias)
    .set_attr_strides({ 1, 1, 1, 1 })
    .set_attr_pads({ 0, 0, 0, 0 })
    .set_attr_dilations({ 1, 1, 1, 1 });

TensorDesc conv2d_input_desc_x(ge::Shape(), FORMAT_NCHW, DT_INT8);
TensorDesc conv2d_input_desc_filter(ge::Shape(), FORMAT_NCHW, DT_INT8);
TensorDesc conv2d_input_desc_bias(ge::Shape(), FORMAT_NCHW, DT_INT32);
TensorDesc conv2d_output_desc_y(ge::Shape(), FORMAT_NCHW, DT_INT32);
conv2d.update_input_desc_x(conv2d_input_desc_x);
conv2d.update_input_desc_filter(conv2d_input_desc_filter);
conv2d.update_input_desc_bias(conv2d_input_desc_bias);
conv2d.update_output_desc_y(conv2d_output_desc_y);

The major steps are as follows:

1. Call the operator type constructor, for example, Conv2D(const string& name) to create an operator instance, and pass the operator name (for example, Conv2d1).

   auto conv2d1 = op::Conv2D("Conv2d")

2. Call set_input_input name to set operator inputs.

       .set_input_x(data)
       .set_input_filter(conv_weight)
       .set_input_bias(conv_bias)

   data is the input node of the entire graph. It is constructed by the Data operator. For details, see Data Operator Definition.

   conv_weight is a constant data constructed by the Const operator. For details, see Const Operator Definition.

   conv_bias is a constant data constructed by the Const operator. For details, see Const Operator Definition.

3. Call set_attr_attribute name to set operator attributes.

   .set_attr_strides({1, 1, 1, 1}) // Set the strides attribute.
   .set_attr_pads({0, 0, 0, 0}) // Set the pads attribute.
   .set_attr_dilations({1, 1, 1, 1}); // Set the dilations attribute.

4. For convolution operators such as Conv2D or operators that are sensitive to C axis processing, you are advised to set the format to NCHW or NHWC in the update_input_desc_input name call. The format must be consistent with the format to be processed.

   TensorDesc conv2d_input_desc_x(ge::Shape(), FORMAT_NCHW, DT_INT8);
   TensorDesc conv2d_input_desc_filter(ge::Shape(), FORMAT_NCHW, DT_INT8);
   TensorDesc conv2d_input_desc_bias(ge::Shape(), FORMAT_NCHW, DT_INT32);
   TensorDesc conv2d_output_desc_y(ge::Shape(), FORMAT_NCHW, DT_INT32);
   conv2d.update_input_desc_x(conv2d_input_desc_x);
   conv2d.update_input_desc_filter(conv2d_input_desc_filter);
   conv2d.update_input_desc_bias(conv2d_input_desc_bias);
   conv2d.update_output_desc_y(conv2d_output_desc_y);