I found a github which implements Tversky-nn and have completed the code to work for NABirds. I also have done quite a bit of perf improvement resulting in each Epoch taking only 13 seconds on my 5090. I believe I can still double the perf but leverage my 2nd 5090 to deal with the computational expensive sample augmentation. This is on my Threadripper 7985WX system with 256GB's of DDR5-6000.
Before I can move to exploring additional use cases I want to duplicate the results in the paper.
Which scheduler and optimizer were used?
When might the code to duplicate the results be open sourced?
I'd like to apply my performance optimizations and explore this further.
I've spent 40 years as a software performance architect.
I found a github which implements Tversky-nn and have completed the code to work for NABirds. I also have done quite a bit of perf improvement resulting in each Epoch taking only 13 seconds on my 5090. I believe I can still double the perf but leverage my 2nd 5090 to deal with the computational expensive sample augmentation. This is on my Threadripper 7985WX system with 256GB's of DDR5-6000.
Before I can move to exploring additional use cases I want to duplicate the results in the paper.
Which scheduler and optimizer were used?
When might the code to duplicate the results be open sourced?
I'd like to apply my performance optimizations and explore this further.
I've spent 40 years as a software performance architect.
🤯