modnet

modnet copied to clipboard

As discussed, it would be nice to show an example of the joint-learning style multifidelity approach

ml-evs

Investigate use of Keras core

1

Keras Core is now in beta https://keras.io/keras_core/, and will replace the current `tf.keras`API in v3. This would allow us to make pytorch, JAX and TF models with the same API,...

ml-evs

I had segmentation faults when performing feature selection on very small datasets (typically for testing). Setting a random seed fixes this issue, and it's always nice to have this feature...

gbrunin

Bump pandas from 1.5.2 to 2.0.1

1

Bumps [pandas](https://github.com/pandas-dev/pandas) from 1.5.2 to 2.0.1. Release notes Sourced from pandas's releases. Pandas 2.0.1 This is a patch release in the 2.0.x series and includes some regression and bug fixes....

dependabot[bot]

I will set up some tests for this in a future PR _Originally posted by @ml-evs in https://github.com/ppdebreuck/modnet/issues/155#issuecomment-1635582379_ This is just a reminder for myself when I get back in...

ml-evs

missing features for prediction

2

If some features are missing (from the optimal ones), when featurizing novel compounds (for predicting stage), `predict()` will not work. This happens for instance when some elements are not present...

ppdebreuck

complex compositions take very long to featurize

4

I would like to run modnet on a dataset in which I have compositions that have very complex stoichiometries. On example would be `C100H3815Br21I279N2185Pb100` To reproduce, this could be an...

Pepe-Marquez

CompositionOnly featurizer preset does not use oxidation states

6

As above. We should probably add the oxidation state featurizers used in the `DeBreuck2020Featurizer` to the `CompositionOnly` featurizer too.

ml-evs

See https://github.com/ppdebreuck/modnet/blob/1fbf7b2f45aee5970ee3f5c6fd88461faefdbc65/modnet/preprocessing.py#L319 Not used at this stage but may be integrated to get RR score. _Reported by @gbrunin_

ppdebreuck

The [feature selection](https://github.com/ppdebreuck/modnet/blob/1fbf7b2f45aee5970ee3f5c6fd88461faefdbc65/modnet/preprocessing.py#L770) procedure relies on `mutual_info_regression` & `mutual_info_classif`which is stochastic. Therefore, feature selection is nondeterministic (the ranked list of optimal descriptors can have slight changes from run to run)....

ppdebreuck