ricomnl

Thanks @abhi18av ! Feel free to reach out if I can help in any way regarding the use case description etc.

I've implemented a working version of this [paper](https://www.biorxiv.org/content/10.1101/2022.01.14.476312v4.full.pdf) in an external scvi module: https://github.com/ricomnl/scvi-ar/pull/1/files. Am currently running some tests but I will create a PR to merge this in

hi @thegregyang , I ran into the same issue and its also due to the depths not matching. In the paper you showed that this method also works for scaling...

@dakoner @mattrasmus this [line](https://github.com/insitro/redun/blob/k8s-rasmus/redun/executors/k8s_utils.py#L180) needs to be changed to: ```python container = client.V1Container(name=name, image=image, args=command, env=env) ``` In kubernetes, the `command` overwrites Dockers `ENTRYPOINT` while `args` overwrites `CMD`. I ran...

Nevermind, I jsut realized it caches everything. Still nice to have the error handling though

it was a fault on my end, the pdf was empty for some reason

I also realized the search is quite slow for 1000s of PDFs. Is this because I'm using a relatively big model or just because they're in PDF format? Would it...

@thegregyang @edwardjhu I saw in the paper in section H, the following: > Of course, in both scenarios, depth, batch size, and sequence lengths can be scaled up and down...

ah yes, so I tried that and it somewhat worked but the result was the below: 4 layers < 8 < 32 < 16