William Patton

There is a little `ArraySource` in the `tests/helpers.py`. Fairly simple to make with any `gunpowder` `Array` class that can be built from numpy arrays and I think most things that...

Regarding the simple augment and the reversed channels, the `SimpleAugment` applies both Transpose and Mirror operations randomly. You limit the transpose operation to the last 2 dimensions, but the `SimpleAugment`...

We don't have a source for reading from OME Zarr. Our `ZarrSource` is pretty basic and avoids making any assumptions about the layout of the zarr. The only assumptions that...

> You can then combine them using RandomProvider, which will randomly pick any of the sources to satisfy a request. Ah I may have misunderstood. I thought the Row/Col/Pos related...

> We could do this on the fly by loading in a region and computing whether it meets the qualifications, but I think it may be more efficient to precompute...

Ah yes. That was recently fixed but we have not yet made new release to propagate the changes.

> Is there a way recommended way to draw from multiple pipelines during training? This should be possible. Here's a quick working example: ```python import zarr import gunpowder as gp...

Should not impact performance at all. I think the likelihood of running into any issues is very low and probably not worth worrying about. If you had multiple pipelines that...

Hi @mattersoflight and @Christianfoley, there is minimal built in support for processing channel dimensions since most of the gunpowder operations are focused on spatial queries and augmentations. There are a...

In general the `prepare` method is supposed to only return the dependencies necessary for the node and should not modify the `request` in place. Best practice is just to create...