You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Apr 29, 2021. It is now read-only.
A simple fix could be to replace the nn.HardTanh operator with a custom layer that just uses torch.clamp instead, but there might be some additional code associated with module.layers, since torch.clamp expects a Tensor (i.e. not a Module). Hopefully, torch.onnx.export is smart enough convert a torch.clamp() call into the appropriate Clip operation(s).
Currently, the DS1 graph fails to export to ONNX due to the
HardTanhoperation not being part of the supported ONNX operators.A simple fix could be to replace the
nn.HardTanhoperator with a custom layer that just usestorch.clampinstead, but there might be some additional code associated with module.layers, sincetorch.clampexpects a Tensor (i.e. not a Module). Hopefully,torch.onnx.exportis smart enough convert atorch.clamp()call into the appropriate Clip operation(s).