This should include pre-built normalizing flow models as well as LightningModule for training.
We could have for example an app.models.MAFAffineFlow(torch.nn.Module) and app.models.CouplingAffineFlow(torch.nn.Module) that automatically apply the ZMatrix + Cartesian transformation, and a KLTrainedFlow(lightning.LightningModule) that takes a flow as input and trains it using KL divergence (uni- and bi-directional).
To implement a fault-robust training in multimap TFEP, we might also have to implement a custom Lightning Trainer capable of resuming mid-epoch.
This should include pre-built normalizing flow models as well as
LightningModule
for training.We could have for example an
app.models.MAFAffineFlow(torch.nn.Module)
andapp.models.CouplingAffineFlow(torch.nn.Module)
that automatically apply the ZMatrix + Cartesian transformation, and aKLTrainedFlow(lightning.LightningModule)
that takes a flow as input and trains it using KL divergence (uni- and bi-directional).To implement a fault-robust training in multimap TFEP, we might also have to implement a custom Lightning
Trainer
capable of resuming mid-epoch.