TensorRT10 conversion replaces Sigmoid with ReLU + Identity #46
Description
Hi, thanks for this great work on DepthAnything TensorRT.
I am trying to convert metric DethAnything model to trt but facing incorrect output depth problems.
Description:
When exporting a metric DepthAnything PyTorch model with a Sigmoid activation to TensorRT via ONNX, the conversion replaces Sigmoid with ReLU + Identity. This changes the network behavior because the output is no longer in [0, 1].
Expected behavior:
Sigmoid activations should be preserved in TensorRT conversion, as TensorRT 10 supports SIGMOID natively. The results are visually correct.
Actual behavior:
nn.Sigmoid() is replaced with nn.ReLU() and nn.Identity() in the converted model, leading to incorrect output values.