P-CBLS: Confidence-Aware Paced-Curriculum Learning by Label Smoothing

This repository contains the code for the paper "Confidence-Aware Paced-Curriculum Learning by Label Smoothing". We evaluate our CBLS and P-CBLS on three common computer vision classification datasets: Tiny-ImageNet, CIFAR10, CIFAR100, and four surgical datasets includeing multi-class classification, multi-label classification, semantic segmentation, image captioning tasks.

The datasets can be found here:

Workflow classificaton: M2CAI 2016 Challenge (http://camma.u-strasbg.fr/m2cai2016/index.php/program-challenge/)

Tool Cassification: Instrument Segmentation Challenge 2017 (https://endovissub2018-roboticscenesegmentation.grand-challenge.org/Data/) and 2018 (https://endovissub2017-roboticinstrumentsegmentation.grand-challenge.org/)

CIFAR-100 and CIFAR-100 Dataset (https://www.cs.toronto.edu/~kriz/cifar.html)

Tiny-ImageNet Dataset (https://www.kaggle.com/c/tiny-imagenet)

The detailed README files for each task can be found inside the corresponding code folder.

Here, we take the training commands on Tiny-ImageNet as the example to demonstrate the usage of our approaches.

Command to train models with CBLS approach on Tiny-ImageNet:

cd cv_classification

Our CBLS:

python main.py --mode CBLS --ls_factor 0.38 --use_cls --cls_decay 0.5 --num_epochs 200 --dataset tinyimagenet

Other baselines including Cross-Entropy Loss (CE), Label Smoothing (LS), Online Label Smoothing (OLS) and Disturb Label (DL).

CE

python main.py --mode CE --num_epochs 200 --dataset tinyimagenet

LS

python main.py --mode LS --ls_factor 0.1 --num_epochs 200 --dataset tinyimagenet

OLS:

python main.py --mode OLS --num_epochs 200 --dataset tinyimagenet

DL:

python main.py --mode DL --num_epochs 200 --dataset tinyimagenet

Command to train models with P-CBLS approach on Tiny-ImageNet:

Obtain the optimal temperature value based on the CE baseline.

python test_temperature.py

Obtain the sorted samples sorted by the calibrated confidence value after pass the optimal temperature value

python cal_confidence.py

Our P-CBLS

python main_pcbls.py