Disparity Between Batches as a Signal for Early Stopping

Requirements

We conducted experiments under:

• python 3.5.2
• torch 1.4.0
• torchvision 0.5.0
• cuda 10.2
• jupyter-notebook 6.0.3
• ipython 7.9.0
• 1 Nvidia Titan X Maxwell GPU

Datasets:

• MNIST, CIFAR-10 and CIFAR-100 will be automatically downloaded by running the script.
• MRNet: In order to download this dataset, register here https://stanfordmlgroup.github.io/competitions/mrnet/.

Description of files

• datasets.py: the code to get data loader for MNIST, CIFAR-10 and CIFAR-100 datasets for a given batch size, training set size and level of label noise.
• dataloader_mrnet.py: the code to get the data loader for MRNet dataset.
• models.py: the code for neural network configurations that are used and two parameter initialization techniques.
• model_mrnet.py: the code for the configuration used for MRNet dataset.
• experiments.py: the code to train the models and compute gradient disparity in each epoch.
• results.ipynb: the code to plot figures after the execution of experiments.py is finished.
• experiments_mrnet.ipynb: the code to train the model for the MRNet dataset.

Example

To train a ResNet-18 on 12.8 k points of the CIFAR-10 dataset with 0 percent label noise level, batch size=128, for 20 epochs run the following command:

python3 experiments.py --dataset cifar10 --numsamples 12800 --batchsize 128 --corruptprob 0 --numepochs 20 --model resnet18 --filename <filename.data>

The results will be saved in <filename.data>. To plot the figures of each experiment run results.ipynb file while reading the <filename.data> file of your choice.

Access to the paper

You can find the full version of the paper (including appendices) at https://arxiv.org/pdf/2107.06665.pdf and the published version at https://2021.ecmlpkdd.org/wp-content/uploads/2021/07/sub_1075.pdf.

Citation

To cite our work please use:

@article{forouzesh2021disparity,
  title={Disparity Between Batches as a Signal for Early Stopping},
  author={Forouzesh, Mahsa and Thiran, Patrick},
  journal={arXiv preprint arXiv:2107.06665},
  year={2021}
}