Yaël Balbastre
Yaël Balbastre
There is almost everything needed in pytorch to write a probabilistic programming (PP) library (see e.g. [Pyro](http://pyro.ai)) and it looks like a good exercise. Plus, if it works well, it...
It's super weird: on github's CI server, the compilation fails [only] on python 3.7. Failure happens at the last linking step (the extension files that binds the c++ code to...
The test that checks that analytical and numerical gradients are identical fails for grid_grad. There must be an error in the c++ code.
NiTorch's c++/cuda library (libnitorch) is compiled against pytorch's c++/cuda library (libtorch). From one version of pytorch to the next, changes in libtorch can cause symbols to disappear, leading to errors...
Currently, the backward passes of grid_push/grid_pull/etc require the forward input tensors as inputs even when the value they hold is not needed. E.g., let the forward pass be `grid_pull(image, grid)`,...
Nitorch imports should use relative indexing, *i.e.*, ```{python} # In file nitorch/bar.py from nitorch.foo import fun # NO from .foo import fun # YES ``` Please reference this issue when...
Sliding boundary conditions are specific to displacement fields: they use a mixture of DCT2 and DST2 boundaries, with the DST2 dimension being different in each channel. Here's an example in...
We need to decide on a license before going public. I'd rather be very permissive. - A solution is GNU-GPL, which is quite permissive but forbids users to distribute closed-source...
- `torch` import all submodules when performing `import nitorch`. That is, we then have access to, _e.g._, `nitorch.nn` without having to import it. This is because `__init__.py` exposes (=imports) all...
PyTorch's convolutions return domain is the joint domain of the input tensor and kernel. This means that the tensor is implicitly cropped by an amount that corresponds to the kernel...