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