[Feature][Relax][ONNX] Support CenterCropPad operator in Relax ONNX frontend #20971

Feature Description

Support the ONNX CenterCropPad operator in the Apache TVM Relax ONNX frontend (python/tvm/relax/frontend/onnx/onnx_frontend.py).

Motivation

CenterCropPad crops and/or pads input tensors centered along specified axes to reach a desired spatial shape. It is a standard building block in computer vision data pipelines and end-to-end vision inference models where image preprocessing is embedded directly into the ONNX graph.

Standardized in ONNX opset 18+, CenterCropPad is currently not supported in the Relax ONNX frontend, resulting in:

NotImplementedError: The following operators are not supported: ['CenterCropPad']

This operator was also highlighted in the ONNX operator coverage RFC tracking issue #20124 under Cropping / Sequence operators.

Operator Specification (ONNX Opset 18+)

• Inputs:
  1. input_data: N-D input tensor (float32, float16, uint8, etc.).
  2. shape: 1D integer tensor specifying the target shape along the specified axes.
• Attributes:
  • axes: list of ints (optional) - the axes along which to crop/pad (defaults to [-2, -1] for spatial dims).
• Outputs:
  1. output_data: Cropped and/or padded tensor matching the target shape.

Expected Behavior

The Relax ONNX frontend should implement CenterCropPad by comparing input dimensions with target dimensions:

1. For dimensions where target < input: compute symmetric start/end indices and apply relax.op.strided_slice.
2. For dimensions where target > input: compute symmetric left/right pad widths and apply relax.op.pad.