FIERCE-repo

This repo contains the code for FIERCE, as presented in our paper Preserving Fine-Grain Feature Information in Classification via Entropic Regularizatio (https://arxiv.org/abs/2208.03684). We also provide the hyperspectral dataset that we used in Section 5.2. Note that we do not provide the dataset age estimation (IMDB & WIKi) and CUB.

Requirements and Installation:

1. A Pytorch installation https://pytorch.org
2. Python version 3.8.1 (lower versions might not work)
3. (optional) Age estimation dataset https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/: IMDB (faces only) and WIKI (faces only)
4. (optional) CUB dataset: https://deepai.org/dataset/cub-200-2011

Datasets

Age estimation and CUB requires to download manually the datasets.

Age Estimation:

it is required to run the following commands to pre-process the dataset.

$ python datasets/process_data_wiki_imdb.py --rootpath 'path/imdb_crop' --metafile 'imdb' --outfile  'imdbfilelist.txt'
$ python datasets/process_data_wiki_imdb.py --rootpath 'path/data/wiki_crop' --metafile 'wiki' --outfile  'wikifilelist.txt'

CUB

We use the split recommanded in https://github.com/icoz69/DeepEMD.

Experiments

Hyperspectral (folder hyperspectral)

• Experiments and ablation studies can be carried out with the notebook: Experiments_Hyperspectral.ipnyb

Age regression (folder age_regression)

• training and saving results: Vanilla

$ python main.py 0  100

FIERCE with parameter 0.3

$ python main.py 0.2 100

Label Smoothing

$ python ls.py 0 100

• plotting the results: with the notebook Stats_age_regression.ipynb. It is possible to plot the results (MSE, prediction) from the file features.pt generated during the training of the model. We provide the features of the 3 models displayed in our paper (Cross-Entropy, Label Smoothing, FIERCE).
• computing mutual information, Stability and Reliability Diagram:

$ python stats.py pathtomodel

models:

• We provide also 3 saved models: model_vanilla.pt (Cross-Entropy Criterion), model_fierce_0.2.pt (FIERCE with parameter 0.2), model_ls.pt (label smoothing) plotting results & statistics:

Few Shot Experiments

• training CIFARFS and print statistics Vanilla

$ python main.py --dataset cifarfs --model resnet18 --lr 0.1 --skip-epochs 300 --entropy 0 --runs 10

FIERCE with parameter 1.5

$ python main.py --dataset cifarfs --model resnet18 --lr 0.1 --skip-epochs 300 --entropy 1.5 runs 10

Label Smoothing

$ python main.py --dataset cifarfs --model resnet18 --lr 0.1 --skip-epochs 300 --entropy 0 --label-smothing 0.1 runs 10 

• training CUB and print statistics

$ python main.py --dataset cubfs --runs 10 --entropy-parameter 0 
$ python main.py --dataset cubfs --runs 10 --entropy-parameter 2 
$ python main.py --dataset cubfs --runs 10 --entropy-parameter 0 --label-smoothing 0.1