search

tnc-br / ddf-isoscapes

4 stars 0 forks source link

Train variational models with real data and compare results #85

Closed jmogarrio closed 1 year ago

jmogarrio commented 1 year ago

Variational model here refers to the TF DNN at https://github.com/tnc-br/ddf-isoscapes/tree/main/dnn.

This is one of the alternatives to xgboost, the other being gpr (some work on this is tracked in #61).

Real data refers to the one referred to by 2023_06_13_Results_CNO in the shared drive.

The points of comparison will be xgboost with the same real data, as well as the initial results of xgboost.

jmogarrio commented 1 year ago

If possible, could we clarify 1. which variational models we're referring to, 2. which real data we're using, and 3. what are we comparing the results against?

jmogarrio commented 1 year ago

Updated Issue description to reflect updates of discussion with @VimWizardVersace. Filed #104, #105, #106.

Have to decide how this issue and #61 relate, as it could be that this is either a follow-up or subtask of the other one.

VimWizardVersace commented 1 year ago

I've generated some new models, and the results aren't much better. I've generated isoscapes for Greta to t-test.

I've implemented it to spec based on go/ddf-variational-models. I'd like to try to rule out any bugs with the tensorflow code, so I think a code walkthrough with either. If the model seems correctly configured, we might just need more/higher quality data.

benwulfe commented 1 year ago

This is our tracking bug for variational model work