Age | Commit message (Collapse) | Author | |
---|---|---|---|
2020-05-19 | Test with explicit weights | Marc Glisse | |
and remove duplicated assignment | |||
2020-05-18 | Infer k when we pass the distances to the nearest neighbors | Marc Glisse | |
2020-05-12 | More test | Marc Glisse | |
2020-05-12 | test + reformat | Marc Glisse | |
2020-04-22 | Reduce the probability of failure of test_dtm | Marc Glisse | |
It is expected that hnsw sometimes misses one neighbor, which has an impact on the DTM, especially if the number of neighbors considered is low. | |||
2020-04-19 | Drop redundant test | Marc Glisse | |
torch.isnan(None) raises an exception anyway | |||
2020-04-14 | Check that the gradient is not NaN | Marc Glisse | |
This can easily happen with pytorch, and there is special code to avoid it. | |||
2020-04-14 | More testing | Marc Glisse | |
2020-04-13 | enable_autodiff with keops | Marc Glisse | |
This doesn't seem like the best way to handle it, we may want to handle it like a wrapper that gets the indices from knn (whatever backend) and then computes the distances. | |||
2020-04-12 | Parallelize the "precomputed" case of knn | Marc Glisse | |
It is supposed to be possible to compile numpy with openmp, but it looks like it isn't done in any of the usual packages. It may be possible to refactor that code so there is less redundancy. | |||
2020-04-11 | Use longer names | Marc Glisse | |
2020-03-28 | Default param of 2 for DTM | Marc Glisse | |
2020-03-28 | Fix test | Marc Glisse | |
2020-03-28 | Fix test | Marc Glisse | |
2020-03-27 | doc | Marc Glisse | |
2020-03-27 | Improve coverage | Marc Glisse | |
2020-03-26 | cmake | Marc Glisse | |