/LeRaC

LeRac: Learning Rate Curriculum

Primary LanguagePython

Learning Rate Curriculum (LeRaC) - IJCV 2024 (official repository)

Introduction

This repository contains the code that implements the experiments described in the work "Learning Rate Curriculum" accepted at the International Journal of Computer Vision.

Citation

Please cite our work if you use any material released in this repository.

@article{Croitoru-IJCV-2024,
  author    = {Croitoru, Florinel-Alin and Ristea, Nicolae-Catalin and Ionescu, Radu Tudor and Sebe, Nicu},
  title     = "{Learning Rate Curriculum}",
  journal = {International Journal of Computer Vision},
  year      = {2024},
  }

Data

The data sets (CIFAR-10, CIFAR-100, Tiny ImageNet, Qnli, BoolQ, RTE) should be stored in a directory called "data":

  | LeRac
    | data
      | CIFAR-10
      | CIFAR-100
      | tiny-imagenet-200
      | imagenet
      | QNLIv2
      | BoolQ
      | RTE

Run

The experiments can be run via a command line, passing the model and the data set as arguments. For example, the command to run the LeRac strategy on Resnet-18 for CIFAR-10, is the following:

    python main.py --model_name resnet18 --dataset cifar10

Obs.:

  1. The process will save the models on disk in a directory called "saved_models". Therefore, this directory should exist before running the experiments.

  2. If the experiments involve a pre-trained architecture, the weights should be stored in a directory called "pretrained_models". Example:

  | LeRac
    | pretrained_models
      | CvT-13-224x224-IN-1k.pth