DeformableConv2D

Description

Deformation convolution.

Input

Four inputs:

x: 4Dtensor in NCHW format.

filter: weight tensor in NCHW format.

offsets: 4-dimensional tensor in NCHW format.

bias: (optional) bias, 1D tensor in ND format.

Output

One output:

y: output tensor in NCHW format.

Property

auto_pad: (optional) supports VALID and NOTSET.

dilations: list of four integers, specifying the dilation rate for dilation convolution. The value range of the H and W dimensions is [1, 255].

groups: number of blocked connections from the input channel to the output channel. Both the input and output channels must be exactly divided by groups. The data type is int32 and must be set to 1.

pads: list of four integers, specifying the top, bottom, left, and right padding. The value range is [0, 255].

strides: list of four integers, specifying the strides of the convolution along the height and width. The value range of the H and W dimensions is [1, 63]. By default, the N and C dimensions are set to 1.

data_format: string, input data format. The default value is NHWC.

deformable_groups: number of group convolution channels. The default value is 1.

modulated: bool, DeformableConv2D version. true indicates v2 and false indicates v1. Currently, only true is supported.

Constraints

Input tensor. The value range of the W dimension is [1, 4096/filter_width]. The value range of H is [1, 100000 / filter_height].

Weight tensor. The value range of the W dimension is [1, 63], and the value range of the H dimension is [1, 63].

If the atc tool does not support the --precision_mode=must_keep_origin_dtype parameter, the input type is float or float64.

ONNX Opset Support

Opset v9/v10/v11/v12/v13/v14/v15/v16