fayolle / pyfrep Goto Github PK

In the module reinit, rename ImplicitNN into ImplicitNNFun.
We will use ImplicitNN for neural nets used for fitting a point cloud.

When creating tensors, check that they have consistent types.
It seems that sometimes we use float32, other times float64. It doesn't always mix well.

Standardize naming conventions and parameters across SDF and non-SDF functions.
For example, we have primitives.cylX() and sdf_primitives.cylinderX().

Use a consistent naming. For example in IO, some of the functions are called 'save' others are called 'write'.

In

Same problem here:

Currently only gyroids and gyroid sheets are implemented.
Add other examples (for example lidinoids)

Add function to export a sampled grid to the VTK file format.

Several SDF primitives are missing (platonic solids, cone variants, triangle, ...). Add them.

Ideally it should be possible to have multiple backends (PyTorch, Jax) and to be able to switch easily between them.
Currently only PyTorch is supported.

Add code to save a surface mesh with a scalar value per node as a VTK file.

Check for missing primitives and operations from the literature on implicit surface and SDF based modeling.

Add an implementation of regularized evolution for fitting parametric primitives.

Add setup.py such that we do not need to add

sys.path.append(os.path.join(os.path.dirname(__file__), "../"))

in the shape scripts.

Change the project name to something else. The problem is that the term "differentiable" does not appear anywhere (the current name is "PyFRep").
Possible names: Diff-FRep, diff-FRep, Diff_FRep, diff_FRep?

Do not forget to change the project name in the file setup.py as well.

Add a simple implementation of slicing.

Add normalization schemes: Rvachev, Taubin (first order only)

The directory 'Examples' should only contain examples showcasing use-cases of the library.
Some of the examples can be removed (for example, one example of micro-structure is sufficient, the conv example is not needed, the gyroid example is probably not needed as well, ...)

Allow to fit a neural net (MLP) to a point cloud or a mesh

Trap NaN in grad, div, ...
It can happen when using r-functions. Use torch.nan_to_num