axeldavy

To the best of my knowledge, this is still an issue. Yours.

> I'd be opposed to this. Cython's basic model is "Python-like syntax and behaviour but using C typing". This is moving away from the Python-like behaviour. I think if you...

I think if you don't declare with the static variable in C++, it's per-instance-per-thread. That's the usage I meant.

From a quick look at the PR, it looks clean, but the binary dll should probably not be part of the commit. @dhewg Thanks for all the work you have...

A quick look at the trace, I see two possible issues: . The game attemps to create D3DFMT_P8 surfaces in system memory and fails (which is probably fine since supporting...

As a following note for my comment on https://github.com/tensorflow/addons/pull/2196#issuecomment-884275962 I've written my own tensorflow addon that does exactly the proposed method. A lot of work remains because I didn't write...

Here is the code I'm using: https://github.com/axeldavy/tensorflow_addons/tree/add-deformable-interpolation The CUDA code can be seen here: https://github.com/axeldavy/tensorflow_addons/blob/add-deformable-interpolation/tensorflow_addons/custom_ops/layers/cc/kernels/deformable_interpolation_op_gpu.cu.cc I don't know enough of XLA to know if it can be efficiently expressed with...

It's not just about training with tensorflow though. One issue with deformable convolution is getting it to work with inference solutions (tensorrt, etc). Ideally there should be an onnx operation...

Your compilation failure doesn't seem related to my changes. I would suggest to just apply my last commit to a branch that works for you.

I hadn't noticed, but onnx has recently received a new operator GridSample that is relevant to the topic. https://github.com/onnx/onnx/blob/main/docs/Operators.md I think it is relevant, because it is better if the...