Add ModelMerger class - Githubissues

muammar / ml4chem

ML4Chem: Machine Learning for Chemistry and Materials

https://ml4chem.dev

Other

93 stars 15 forks source link

Add ModelMerger class #10

Closed muammar closed 4 years ago

muammar commented 5 years ago

[x] Model merger.
[x] Loss function weighted sum vs independently.
[x] Printing of training evolution.
[x] Save model (I think code already available in ml4chem could be used for that).
[x] Set convergence stop mechanism.
[x] Parallelization.
[x] Add more compatibility with different available components.
[x] Check for consistency of the class (independent vs dependent loss).

That class would help to merge models like this:

Chgpo

muammar commented 5 years ago

This class is now working for cases where the loss functions of models are independent (backward propagation occurs independently as well), and when they are summed (weights of all models are updated according to the gradient of all of them depending on each other).

The parallelization has to be worked out because a new forward() method was created.

muammar commented 5 years ago

The parallelization has been worked out because a new forward() method was added. I just tested with MSE loss functions. I also noticed that the independent loss function case is broken now.