Roman Feldbauer

The dtype mismatch still seems to be an issue with `metric='precomputed'`. I get the error on fitting an `np.float32` distance matrix, but casting to `np.float64` fixes the problem.

Not yet, this just caught my eye while reading the code. I'll have a closer look and actually implement & benchmark this.

On a closer look `UnetModel.eval()` seems broken. For example, there are undefined variables used [here](https://github.com/MouseLand/cellpose/blob/8a5360a7b7db7472d5dd21bd2cd5d050e421fea5/cellpose/core.py#L253) (channel_axis, z_axis), and [here](https://github.com/MouseLand/cellpose/blob/8a5360a7b7db7472d5dd21bd2cd5d050e421fea5/cellpose/core.py#L319) (nolist). That is, the function raises a NameError. I suppose, vanilla...

Imo, the benefit would be reduced maintenance effort in cellpose development. The optimization proposal (2) would be obsolete then. Users will anyway see no difference. Proposal (1) about segmentation of...

IIRC, nmslib's HNSW does not support any metric besides Eucl and cos, but please feel free to point me to documentation that states otherwise. However, this code seems to fail...

Indeed, while optimized indices are only available for Eucl and cos, many more spaces are supported in general. For personal reference, the detailed list on supported spaces is available in...

It seems ONNG can be enabled in ngtpy, but it is currently not documented. However, there is an example here: https://github.com/yahoojapan/NGT/issues/30

New NGT release 1.7.10 should fix this: https://github.com/yahoojapan/NGT/releases/tag/v1.7.10

1.8.0 brought docs for ONNG. It is already activate here, but index building is extremely slow due to difficult parameterization. Need to check.

Thanks for your interest. That's something I've been thinking about, but never found time to actually check. BERT embeddings are typically high-dimensional, so hubness might play a role. You could...