Discrepancy between PyTorch and ONNX depth anything v2 small models

[dfatimazohra]

HelloI'm seeing a discrepancy between PyTorch and ONNX depth anything v2 small models.

Does anyone have any solution for this issue?
Thanks

[fabio-sim]

Hi @dfatimazohra, thank you for your interest in Depth-Anything-ONNX.

Due to differing implementations, the outputs of torch and ORT usually have slight differences on the order of 1e-3 ~ 1e-4. If you're observing something much larger (e.g., on the order of 1e-1), could you provide the details (what inputs, which models) to reproduce it?