Latent dimensionality consistency

[toushi68]

Hi,
On the way to finger out the reason of the very low accuracy, I noticed there are some inconsistencies. For the grammar based training, in train_zinc.py, the default LATENT = 56, but in model_zinc.py, latent_rep_size = 2 appeared twice: in def create(....) and def load(....)
For the string based training,, in train_zinc_str.py, the default is also LATENT = 56, but in model_zinc_str.py, latent_rep_size = 292 appeared twice. in def create(....) and def load(....).
Is it normal? Is it necessary to keep them consistent for each type of the training?
In keras-molecule, the default latent is 292, for your grammar based latent space, Is there any particular consideration to make it 56?
Thanks!
Toushi68

[mkusner]

Hey Toushi68,

Sorry for the confusion. We used a latent dimensionality of 56 in all of our experiments, except when we visualize the latent space (Figure 6 in the appendix: https://arxiv.org/pdf/1703.01925.pdf), where we use a dimensionality of 2.

The reason we chose 56 was because this was used in the CVAE paper: https://pubs.acs.org/doi/full/10.1021/acscentsci.7b00572 I believe they found this number using Bayesian optimization to tune this hyperparameter.

So my recommendation is to always keep it 56 dimensional (unless you do more tuning :))).

[toushi68]

thank you for the clarification ! Toushi68 mkusner <notifications@github.com> 于 2018年10月15日周一 06:40写道：

…

Hey Toushi68, Sorry for the confusion. We used a latent dimensionality of 56 in all of our experiments, except when we visualize the latent space (Figure 6 in the appendix: https://arxiv.org/pdf/1703.01925.pdf), where we use a dimensionality of 2. The reason we chose 56 was because this was used in the CVAE paper: https://pubs.acs.org/doi/full/10.1021/acscentsci.7b00572 I believe they found this number using Bayesian optimization to tune this hyperparameter. So my recommendation is to always keep it 56 dimensional (unless you do more tuning :))). — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#14 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AiR0o_xDW1e7LNMgUQFAZymuy9BbwC9Pks5ulGYNgaJpZM4VW-Y_> .