ATEN | Qixian Zhou , Xiaodan Liang

Have you tried to export after defining the operator for onnx? Something along the lines of the following code by Huawei.

On another note, when loading a model, you can technically override anything you want. Putting a specific layer to equal your modified class that inherits the original, keeps the same behavior (input and output) but execution of it can be modified. You can try to use this to save the model with changed problematic operators, transform it in onnx, and fine tune in such form (or even in pytorch).

This generally seems best solved by the onnx team, so long term solution might be to post a request for that specific operator on the github issues page (but probably slow).

You're observation is correct, in order to perform the update you should:

1. Apply the modification with in-place operators.

2. Wrap the calls with torch.no_grad context manager.

For instance:

Turns out I had to change the function to built the tensor, found at https://discuss.pytorch.org/t/unknown-builtin-op-aten-tensor/62389

I found that this works:

auto& state = static_cast(*state_[c10::guts::to_string(p.unsafeGetTensorImpl())]);

very simple and juicy!

ValueError: Unsupported ONNX opset version N -> install latest PyTorch.

Credit to Tianleiwu on this Git Issue.

As per 1st cell of Notebook:

The problem is that you try to fill in a bool Tensor which is apparently not yet supported in jit (or a bug)

Replacing this:

You might use the DataParallel or DistributedDataParallel framework in Pytorch

Maybe the problem comes from this line:

torch.backends.cudnn.enabled = False

You might comment or remove it and try again.

AVX, AVX2, and FMA are CPU instruction sets and are not related to multi-threading. If the pip package for pytorch/caffe2 used these instructions on a CPU that didn't support them, the software wouldnt work. Pytorch, installed via pip comes with multi-threading enabled though. You can confirm this with torch.__config__.parallel_info()