Describe the feature and motivation

In an attention subgraph, there are three Slice operators. Slice operator can have up to 5 inputs: data, starts and ends are required, while axes and steps are optional. In the current dnn graph simplifier, the number of inputs must match what we hardcoded in the code. If we want to support all Slice patterns, we need to make several duplicates whose only difference is the number of inputs of Slice operator.

Example:

• VIT_B_16 (Slice have 4 inputs, axes is default value "-1"; See below for the code exporting model)
• VitTrack (Slice have 5 inputs, steps is default value "1"): https://github.com/opencv/opencv_zoo/blob/main/models/object_tracking_vittrack/object_tracking_vittrack_2023sep.onnx

Current workaround is a manual modification on VitTrack which removes the steps input.

Additional context

Code to export VIT_B_16:

import torch # version: 1.13.1
import torchvision # version: 0.14.1
from onnxsim import simplify # version: 0.4.33

dummy_input = torch.randn(1, 3, 224, 224, device="cpu")
input_names = [ "input" ]
output_names = [ "output" ]

opset = 17

# get and export model
model = torchvision.models.vit_b_16(weights=torchvision.models.ViT_B_16_Weights.DEFAULT)
torch.onnx.export(model, dummy_input, "vit_b_16.opset{}.onnx".format(opset), verbose=False, export_params=True,
                  opset_version=opset, input_names=input_names, output_names=output_names)
# simplify model
model_simplified, check = simplify("vit_b_16.opset{}.onnx".format(opset))
onnx.save(model_simplified, "vit_b_16.opset{}.sim.onnx".format(opset))