/Implicit_geometry

A Repository to recreate the results of "On the Implicit Geometry of Cross-Entropy Parameterizations for Label-Imbalanced Data"

Primary LanguagePython

Implicit_Geom_CE_Param_Imbalance_Data

A Repository to recreate the results of "On the Implicit Geometry of Cross-Entropy Parameterizations for Label-Imbalanced Data" .

Reproduce Experimental Vs Theory Comparison

plot plot

All scripts required to reproduce theory-vs-exp are provided in train_models. As an example, in order to produce the above results for CDT: \

CIFAR10 + ResNet18

python main_deepnet.py --gpu --loss_type CDT --model ResNet18 --dataset CIFAR10

MLP + ResNet18

python main_deepnet.py --gpu --loss_type CDT --model ResNet18 --dataset CIFAR10

UFM

python main_UFM.py --loss_type CDT

The above commands will perform the experiments along a range of $\gamma \in [ -1.5, -1.25, ..., -0.25, 0.0, ..., 1.5 ]$ for $ R = 10 $ step imbalance ratio for one iteration (without data Augmentation). MNIST and CIFAR10 datasets will be downloaded into .\data folder. Results will be saved into proper directories in .\saved_logs as log.pkl files.

In order to produce the plots, run :

python geom_compare.py --loss_type CDT

Same resutls can be reproduced for LDT.