Updated FeatureClassifier: partial update, still integrate to 1e4 #40

Ethadhani · 2024-12-23T19:35:23Z

Made multiple changes that only affect FeatureClassifier

Integration is still to 1e4 orbits of inner most planet

Changed implementation of how features are generated
Ensured model compatibility with new version of REBOUND
Re-trained model with cleaned data and tested impact of duplications/mislabeling in original dataset
Removed systems that go unstable in short integration from testing dataset
Updated saved model

Change comparisons, tests, and backwards compatibility

In paper_plots

testUnstable.ipynb demonstrates change in AUC calculation to not include systems that go unstable in short integration
cleanVnonClean.ipynb compares model trained with clean data verses not clean data.
oldReboundNewRebound.ipynb compares old rebound verses new rebound

commit 2a79a42 contains additional info and comparisons of:

Old REBOUND to new REBOUND with training/testing
Model trained with clean verses dirty data
Initial Tsec tests

commit 9468319 contains comparisons for old feature generation implementation verses new implementation.

commit 086be72 contains all updates except for Tsec related integration and features, this version is avalable for backwards compatibility.

Data filtration/clean data generation is done in accordance to, and with the data from, 10.5281/zenodo.14327471

…arison. This new method allows more flexibility for future advancements/new features without impacting the deep regressor, while decreasing computation time.

…d imports

…dentical features, proper documentation to come

…e and full examples that the old feature generation method and new feature generation method yield the exact same results and are compatible to each other

…od, also isolated featureclassifier from other models such that featureclassifier.py, ClassifierSeries.py, and features.py operate together and independently from the rest of the spock system to allow changes to be made

…n since sim.copy() is not supported

… make a copy of simulation if using new rebound, this is because old rebound does not support sim.copy()

…n. Was supposed to state, fixed pool function to generate training data and removed redundant file, made it so that the training data pool runs through jupyter notebook

…fically accounted for rebound.Collision and rebound.Ejected

…example

…ameters, removed old file

…earn 1.7 is unstable

fixed depreciating warning by updating used XGBoost version, this also decreases training time and increases performance by marginal amount. (was using 1.7, now 2.1)

…n error

…viate sklearn error

… match old docs

gitnotebooks · 2024-12-23T19:35:26Z

Found 11 changed notebooks. Review the changes at https://app.gitnotebooks.com/dtamayo/spock/pull/40

Ethadhani added 30 commits July 10, 2024 22:25

Merged new feature series method, as well as temp model file for comp…

9468319

…arison. This new method allows more flexibility for future advancements/new features without impacting the deep regressor, while decreasing computation time.

fixed discrepancy between strongest MMR calculations, as well as fixe…

82bd82e

…d imports

update __init__ to include alternative featureKclassifier method

0118284

Made old and new version of spock entirely compatible and producing i…

fce5dd2

…dentical features, proper documentation to come

check point, will try implementing multiprocessing

c8be93c

fixed multi processing and increased speed

b9b771d

removed redundant code

31bab28

fixed num of data collections

6d9e700

Worked on comparison pipeline

79e990b

Cleaned up comments

23e0ca8

PIPELINE: CompairFeatGen.... and compairHelper.py contain the pipelin…

1aee18d

…e and full examples that the old feature generation method and new feature generation method yield the exact same results and are compatible to each other

remove old feature classifier file, this version is non functional

155561c

remove comparison files of old and new method

180bb51

new train doc

68b8c9c

New training data generation process

d093f3b

remove empty file

d07061f

Started working on new model training

16e46b9

created new spock training file, and removed old notebooks

b44a675

rename training doc

8251f23

added source of data

44f7164

added old rebound method

8968144

added support for old rebound megno

419d212

removed redundant file and fixed pool

5888138

made spock change the simulation in place if using old rebound versio…

108a263

…n since sim.copy() is not supported

made changes to change simulation in place if using old rebound, will…

9159c20

… make a copy of simulation if using new rebound, this is because old rebound does not support sim.copy()

not the last commit but the one before that, has the wrong descriptio…

1226c82

…n. Was supposed to state, fixed pool function to generate training data and removed redundant file, made it so that the training data pool runs through jupyter notebook

made spock data feature generation compatible with old rebound, speci…

c69cc4d

…fically accounted for rebound.Collision and rebound.Ejected

fixed try except to handle unstable system

a9ccb8c

Generated old rebound training data

e6df5ff

Ethadhani and others added 28 commits November 27, 2024 15:56

removed old models

7634afa

re trained and updated model with Tsec integration and feature

6d09da3

Example of generating data using old rebound for pre cleaned data as …

53804f1

…example

comparison of testing only on systems that are stable after short int

8f45664

clean data verses not clean data comparison

d681815

updated rebound comparison

a512504

Tsec Ablation study

1ec64c7

remove old rebound dirty data generation example

5c1b193

Made comparisons print AUC and FPR

12836b5

re trained and updated model in accordance to original paper hyperpar…

20a0f74

…ameters, removed old file

small style change

2746c81

Update notebook examples

7586b9a

Updated example notebooks

9757fec

fixed print typo and re ran

ae2bcf3

reviewed comments, updating documentation and style

deccbdf

updated scikit-learn requirment due to depreciation warning, scikit-l…

b73b36f

…earn 1.7 is unstable

un changed sklearn dependency

fc68aad

updated model trained with XGBoost version 2.1

dad1d29

fixed depreciating warning by updating used XGBoost version, this also decreases training time and increases performance by marginal amount. (was using 1.7, now 2.1)

reran all examples using updated XGBoost version

5713f12

deleted model file

790bd3b

re trained model using unstable XGBoost 3.0.0 to try alleviate sklear…

f1210e1

…n error

test: upload retrained model using unstable XGBoost 3.0.0 to try alle…

c103aa1

…viate sklearn error

re trained model using older (~2.1) XGBoost and changed saved name to…

5c1cde4

… match old docs

changed scikit-learn requirement until xgboost bug is patched

d92153e

SPOCK FC version with fully integrated Tsec

ccde5ab

fixed dependency

a22d7b2

SPOCK FC version with fully integrated Tsec

5c4e548

SPOCK 1e4 integrated version

086be72

Ethadhani closed this Dec 24, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Updated FeatureClassifier: partial update, still integrate to 1e4 #40

Updated FeatureClassifier: partial update, still integrate to 1e4 #40

Ethadhani commented Dec 23, 2024

gitnotebooks bot commented Dec 23, 2024

Updated FeatureClassifier: partial update, still integrate to 1e4 #40

Updated FeatureClassifier: partial update, still integrate to 1e4 #40

Conversation

Ethadhani commented Dec 23, 2024

Made multiple changes that only affect FeatureClassifier

Integration is still to 1e4 orbits of inner most planet

Change comparisons, tests, and backwards compatibility

In paper_plots

gitnotebooks bot commented Dec 23, 2024