Please report any issue through the issue tool of Github and not through email.

• arXiv: https://arxiv.org/abs/1911.10720

• If you use this code, please cite our work:

@article{belharbi2019unimoconstraints,
  title={Non-parametric Uni-modality Constraints for Deep Ordinal Classification},
  author={Belharbi, S. and Ben Ayed, I. and McCaffrey, L. and Granger, E.},
  journal={coRR},
  volume={abs/1911.10720},
  year={2019}
}

• Method overview
• Demo
• Results
• Requirements
• How to run the code
• Reproducibility
• MultiGPU support and reproducibility
• Synchronized-Batch-Norm support (MultiGPUs)

See ./cmds on how to download the datasets. You find the splits in ./folds. Unzip first (folds.zip). The code that generated the splits is ./create_folds.py.

We use Pytorch 1.2.0 and Python 3.7.0. For installation, see ./dependencies for a way on how to install the requirements within a virtual environment.

python main.py --cudaid your_cuda_id --yaml basename_your_yaml_file

You can override the values of the yaml file using command line:

python main.py --cudaid 1 --yaml bach-part-a-2018.yaml --bsize 8 --lr 0.001 --wdecay 1e-05 --momentum 0.9 --epoch 4 --stepsize 100 --modelname resnet18 --alpha 0.6 --kmax 0.1 --kmin 0.1 --dout 0.0 --modalities 5 --pretrained True  --dataset bach-part-a-2018 --split 0 --fold 0  --loss LossCE

See all the keys that you can override using the command line in tools.get_yaml_args(). You can generate the yaml files using ./yaml-gen.py

• All the experiments, splits generation were achieved using seed 0. See ./create_folds.py
• All the results in the paper were obtained using one GPU with 16GB (not really used) of memory and less than 10GB of RAM.
• Please report any issue with reproducibility.

You need to to set the paths to the data on your own. An error will be raised when the host is not recognized. See the function in tools.get_rootpath_2_dataset().

We took a particular care to the reproducibility of the code.

• The code is reproducible under Pytorch reproducibility terms.

Completely reproducible results are not guaranteed across PyTorch releases, individual commits or different platforms. Furthermore, results need not be reproducible between CPU and GPU executions, even when using identical seeds.

• The code is guaranteed to be reproducible over the same device INDEPENDENTLY OF THE NUMBER OF WORKERS (>= 0). You have to use a seed in order to obtain the same results over two identical runs. See ./reproducibility.py
• Samples can be preloaded in memory to avoid disc access. See ./loader.PhotoDataset(). DEFAULT SEED IS 0 which we used in all our experiments.
• Samples can be preloaded AND preprocessed AND saved in memory (for inference, i.e., test). However, for large dataset, and when the pre-processing aims at increasing the size of the data (for instance, upsampling), this is to be avoided. See ./loader.PhotoDataset()
• We decorated sensitive operations that use random generators by a fixed seed. This allowed more flexibility in terms of reproducibility. For instance, if you decide to switch off pre-processing samples of a dataset (given that such operation relies heavily on random generators) and do the processing on the fly, such a decorator seed allows the state of the random generators in the following code to be independent of the results of this switch. This is a work-around. In the future, we consider more clean, and easy way to make the operations that depend on random generator independent of each other.

• The code supports MultiGPUs.
• Despite our effort to make the code reproducible in the case of multigpu, we achieve reproducibility but, sometimes it breaks. See this.

The code supports synchronized BN. By default, SyncBN is allowed as long as multigpu mode is on. You can prevent using SynchBN, and get back to the standard Pytroch non-synchronized BN, using bash command, before running the code:

# $ export ACTIVATE_SYNC_BN="True"   ----> Activate the SynchBN
# $ export ACTIVATE_SYNC_BN="False"   ----> Deactivate the SynchBN

All the credits of the SynchBN go to Tamaki Kojima([email protected]) (https://github.com/tamakoji/pytorch-syncbn).