Closed kiramt closed 3 years ago
Hi Kira, thanks for trying fastISM out! You're right, it seems to be bugging out for a batch size of 1. I'll look into it.
fastISM runs optimally when GPU memory is maxed out and is run on the most sequences possible in a batch. For small batch sizes it is quite possible it would end up being slower than a standard implementation (due to overheads). If you could describe your use case roughly I may be able to offer more help.
Thanks Surag. I'd expect mostly I'd be running with larger batch sizes anyway, and could fall back on the standard implementation if a small batch was required. I was using the single sequence (with my own model etc) just as a check that I had my input and output processing set up correctly, so then went back to the tutorial when I was getting an error to see if I had done something wrong.
Sounds good! Please don't hesitate to reach out if you get stuck. I'll get to the batch size 1 case soon.
Hi Surag, I also get the same error if I input 2 sequences which are not identical e.g. if I set chr3_enhancer_a to chr3_enhancer but with the first base set to G instead, and have sequences = [chr3_enhancer, chr3_enhancer_a].
Hi Kira, I've pushed some fixes to v0.4.2
. Please give it a try and let me know if it works. Thanks!
Thanks Surag - that seems to have fixed it!
Great, thanks!
Hi - I saw your MLCB talk and was hoping to try out fastism. I was working through the code in your tutorial, as below:
It runs fine when I supply 5 x chr3_enhancer but if I make it a batch of 1 sequence I get the following error: