florpi/graphcosmologynetwork

Graph Neural Network implementation to study the Galaxy-Halo connection.

graphcosmologynetwork's Issues

add_argument option for label selection

Implement arg parse option to select what to learn: dark_or_light, nr_of_galaxies, central_or_satellite, ...

Feature decomposition

1. Test PCA, LDA, FA, ... in .ipynb and implement in .py

Constant bias in 2PCF

1. Error analysis on the RF outputs and its impact on the 2PCF
2. Check whether same problem with other algorithms like FCN
3. Jacknife errorbars

Graph issues

1. check the possibility of giving different weights to the main-node and its neighbours., effect of environment could be small(er).
2. think about different quantity to assign neighbour distances since 1/distance^2 can get quite small (also has units, maybe gravitational force).

Sacred & TensorBoard for Networks

Rewrite models (sklearn and pytorch, order according to priority) to utilize sacred and tensorboard for management of experiments.

[GCN] Test scipy.spatial.cKDTree().query()

Run @pytest.mark.parametrize("particular_edge", [0, 100, 2000, 100000]) on new scipy.spatial.cKDTree().query() graph construction.

Move RandomForest_transition from notebook to script

As the title says plus create new train.py file for sklearn and pytorch

Implement SMOTE of imblearn in dataset balance

For dataset interpolation use implement from imblearn.over_sampling import SMOTE in balance_dataset.py

[RDF/Features] Add new Features and run tests

1. Add new features such as: angular momentum, epoch of formation, environment, central density, Haas environment.
2. Run new tests of RDF and check for correlations and importance.

Feature extraction from position and velocity profiles