Jonathan Fischoff
Jonathan Fischoff
I didn't run the complete training like that. I just did a test. I think it took 2 days on A100 8x
My first attempt was to directly transliterate the paper, which involved lots of direct pointer management and was just too difficult for me. I'm trying a second route that is...
cool. I didn’t know about this library and wrote a similar one that uses a pipe. Those sound like perfectly reasonable requirements. On Fri, Jun 1, 2018 at 4:41 PM...
I appreciate the offer. In all honesty I’m only 50% certain I’ll put a PR up at all ... but yes assuming I do :p On Fri, Jun 1, 2018...
> If a pipe is full, writing to it will block You need to read from it on another thread.
> You have to put it either onto disk or into memory. The goal is to avoid writing to disk and then reading from disk and just buffer in memory....
> I think going via the disk would be beneficial if silently actually offered functions via which the captured output could be read incrementally/streamingly Agree
What is MMD?
> But maybe could just use dicts. Can't we use the output types we define?
Insightface uses multiple models. For instance it is uses mtcnn for initial face detection which is MIT IIRC. What models are being used specifically here?