search

tstandley / taskgrouping

Code for Which Tasks Should Be Learned Together in Multi-task Learning?
Other
93 stars 14 forks source link

Pretrained Models? #2

Open vasusingla opened 4 years ago

vasusingla commented 4 years ago

Hi, and thanks for the great paper! Will you be releasing all the 31 models for all the possible combinations of the 5 tasks? It'll be extremely useful for my research.

tstandley commented 4 years ago

Hey Vasu, Nice to hear that the work is appreciated.

Did you read the version on Arxiv? It's kinda old now. The paper was just accepted to ICML and I'll be doing an update shortly.

In the new version, there are 4 sets of 31 networks (actually 36 if you include the half-sized networks).

The first set of models, which I analyze in the arxiv version, no longer works with the current codebase. I can give them to you, but I won't be releasing them publicly because they'll only cause headaches. You probably want the second/fourth set anyway. I'll be releasing those sometime this month. If you need them faster, let me know exactly which ones you want and I'll get them to you.

Best, Trevor

On Mon, Jul 6, 2020 at 4:10 PM Vasu Singla notifications@github.com wrote:

Hi, and thanks for the great paper! Will you be releasing all the 31 models for all the possible combinations of the 5 tasks? It'll be extremely useful for my research.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tstandley/taskgrouping/issues/2, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEDABXSCTSXMFD6B652F4JTR2JKUPANCNFSM4OSBSH2A .

vasusingla commented 4 years ago

I see I suppose the second/fourth set would be best. I can wait for your release, thanks a lot!

tstandley commented 4 years ago

Pretrained models for the second set are released. Let me know if you have any issues using them. Releasing set 4 seems like it might be trickier. What are you wanting the models for?