DeformableConv2D

Description

Computes a two-dimensional deformable convolution with the given four-dimensional X, filter, and offsets tensors.

Input

• X: 4D input tensor of type float16 or float32, with format NCHW.
• filter: 4D tensor of the learnable filter, which must be of the same type as X. The format can be NCHW and the data type can be float16 or float32.
• offsets: 4D tensor of the x-y coordinate offsets and mask, with format NCHW and data type float16 or float32.
• B (optional): bias, 1D input tensor of type float16 or float32, with shape [M] and format ND.

Attribute

• auto_pad (optional): The value can be VALID or NOTSET.
• dilations: a list of four integers, specifying the dilation rate for dilation convolution. The value range of the H and W dimensions is [1, 255].
• groups (optional): int32. The default value is 1, indicating the number of blocked connections from the input channel to the output channel. Both the input and output channels must be exactly divided by groups.
• pads: list of four integers, specifying the number of pixels to add to each side of the input. The value range is [0, 255].
• strides: a list of four integers, specifying the convolution strides along the height H and width W. The value range for the H and W dimensions is [1, 63]. By default, the N and C dimensions are set to 1.
• data_format: string, specifying the format of the input data. The default value is NHWC.
• deformable_groups: number of channels for group convolution. The default value is 1.
• modulated: bool, specifying the DeformableConv2D version. true indicates v2 and false indicates v1. Currently, only true is supported.

Output

y: a tensor of type float16 or float32, in NCHW format.

Constraints

• For the input tensor X, expected range of the W dimension is [1, 4096/filter_width] and expected range of the H dimension is [1, 100000/filter_height].
• For the weight tensor, expected range of both the W and H dimensions are [1, 63].
• The operator does not support inputs of type float and float64 when the atc command-line option --precision_mode is set to must_keep_origin_dtype.

ONNX Opset Support

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