First implementation of a Variation Autoencoder CV

[andrrizzi]

Description

This implements a CV based on a standard variational autoencoder using a MSE reconstruction loss and regularizing with a KL divergence w.r.t. a Normal(0, 1) distribution. For the CV, only the mean is used as CV in production.

Questions

• I forgot, did we decide to have a mechanism to do the preprocessing inside training_step()? For now, I'm not doing any preprocessing during training.

Status

• Ready to go

[luigibonati]

do you think it makes sense to make the loss function return also the contributions to the loss (reconstruction and KL) and log them?

[luigibonati]

I think that also for the TDA CV @EnricoTrizio might want to keep track of the contributions to the loss, so we might think of a uniform way of doing this

[andrrizzi]

Yeah, I think that's a good idea. Maybe we could return only the loss as default and control whether the decomposition in kl and reconstruction is returned with an optional argument of elbo_gaussians_loss()?

[luigibonati]

yes, that sounds good. i think we just need to think about a general way of returning and logging that works in every case (and also if a custom loss is used). in general, it would be good to return a dict so that the keys determine the log name. actually this would involve also lda/tica CVS (there we want to check the values of the single eigenvalues) so it is pretty general.

Three options come to my mind:

1. we make the loss functions return either a scalar loss or a dictionary that contains the 'loss' as well as additional keys and items to be logged. we then need to parse the variable in the training_step func and check which case we are in
2. we return either a scalar loss or a tuple (loss, log_dict) and do the same as in 1)
3. we change every loss to return not a scalar but always a dict

I would go for option no. 3 which is the cleanest one

what do you think?

[andrrizzi]

I would avoid 2 since it makes logging harder, but both 1 and 3 seems good to me.

If I had to choose, I'd probably go with 1, returning the total loss by default (simply because it's what a user would likely expect when using a loss function object) and have an attribute that controls whether a more detailed dictionary is returned (which can be turned on in the default loss_options in the cvs).

[luigibonati]

Questions

• I forgot, did we decide to have a mechanism to do the preprocessing inside training_step()? For now, I'm not doing any preprocessing during training.

No, the way you implemented is correct and uniform with the other classes (e.g. overloading forward_blocks) in the sense that base cv class has a forward method that adds pre and post processing modules to forward_blocks. As it is now, this is not supposed to take place during training