search

lucpaoli / SAFT_ML

2 stars 0 forks source link

Make association term differentiable (high importance) #18

Open lucpaoli opened 10 months ago

lucpaoli commented 10 months ago

Currently dispchain & hs terms are used in the loss. We really need association for any of this to be worth it.

Problem: Differentiating through the nonlinear solve

longemen3000 commented 10 months ago

If you are only training single components with normal association sites, you should be only hitting the exact solver, that has an analytical expression

lucpaoli commented 10 months ago

Oh you're so right, thank you!