In numba/stencils/stencil.py, there are various places (like line 552, "if isinstance(kernel_size[i][0], int):") where we check for "int" in relation to neighborhoods. I ran across a case where I was creating a neighborhood tuple by extracting values from a Numpy array. This causes a problem because those Numpy values will not match in these isinstance int checks. I worked around it by conver

Problem: the approximate method can still be slow for many trees
catboost version: master
Operating System: ubuntu 18.04
CPU: i9
GPU: RTX2080

Would be good to be able to specify how many trees to use for shapley. The model.predict and prediction_type versions allow this. lgbm/xgb allow this.

Today, I can manipulate ListDtype columns and execute operations like segmented sort and unique. Because these operations do not have cross-row (or cross-partition) dependencies, they can be executed in Dask by passing a lambda function to map_partitions.

It would be nice to expose the list accessor on dask-cuDF objects like we do for other accessors. As this is not supported by pandas, per

Current implementation of join can be improved by performing the operation in a single call to the backend kernel instead of multiple calls.

This is a fairly easy kernel and may be a good issue for someone getting to know CUDA/ArrayFire internals. Ping me if you want additional info.

Problem

Cub allows itself to place into a namespace via CUB_NS_PREFIX and CUB_NS_POSTFIX, such that multiple shared libraries can each utilize their own copy of it (and thus different versions can safely coexist). Static variables used for caching could otherwise cause problems (e.g., https://github.com/NVIDIA/cub/blob/main/cub/util_device.cuh#L212).

Thrust however depends on cub and

Describe the bug
Not a proper bug. The UMAP implementation has a typo in the target_weights argument, where the original UMAP uses target_weight. This creates issues of compatibility when working with both libraries.

Steps/Code to reproduce bug

from cuml import UMAP

mapper = UMAP(n_neighbors=15, n_components=2, target_weights=0.5)
# no error here

mapper2 = UMAP(n_neigh

I often use -v just to see that something is going on, but a progress bar (enabled by default) would serve the same purpose and be more concise.

We can just factor out the code from futhark bench for this.

Thank you for this fantastic work!

Could it be possible the fit_transform() method returns the KL divergence of the run?

Thx!