This is a Pytorch implementation of the model described in our paper:

Q. Xu, Z. Yang, Y. Jiang, X. Cao, Q. Huang, and Y. Yao. Not All Samples are Trustworthy: Towards Deep Robust SVP Prediction. TPAMI 2021.

Dependencies

• PyTorch >= 1.0.0
• torchvision
• numpy
• scipy
• Pillow

Data

The train/val/test split annotation file is in the data/ folder for age/lfw/shoes datasets. Please unzip data/images.zip before training.

For age and lfw, the data format is (file_1, file_2, label, attr_id, pair_id, strength); for shoes, the data format is (attr_id, img1_id, img2_id, pair_id, label, strength).

• file_1/file_2: file name of images
• label: pairwise comparison label (-1/1)
• attr_id: attribute id, starting from 0.
• pair_id: pair id given a specific attribute, starting from 0. Note that (i,j,-1) and (i,j,1) are the same pair but different samples for the same attribute.
• strength: how many times a label is annotated for a pair given a specific attribute.

For custom dataset, you can write a class that inherits from the base class PairwiseDataset in utils.py like we do in LFWDataSet.

Model

We provide four implementations of our model: two sparse-learning-based (LS-Deep-with-gamma and Logit-Deep-with-gamma), and two contraction-loss-based (Huber-Deep and RLogit-Deep) described in the paper.

Train

Here is an example to train the models for lfw dataset.

python train_gamma_lfw.py --lamda=1.2 -e=30 --loss="l2"
python train_gamma_lfw.py --outer=10 --lamda=0.5 -e=5 --loss="logit" -L=20
python train_wo_gamma_lfw.py -l=1e-3 -e=70 --loss="huber" --lamda=0.1
python train_wo_gamma_lfw.py -l=1e-3 -e=70 --loss="rlogit" --lamda=0.1

Hyperparameter:

• lambda (lambda1 in our paper): controls the sparsity of gamma or the contraction margin.
• e: the number of epochs.
• l: learning rate.

Citation

Please cite our paper if you use this code in your own work:

@inproceedings{xu2021not,
  title={Not All Samples are Trustworthy: Towards Deep Robust SVP Prediction},
  author={Xu, Qianqian and Yang, Zhiyong and Jiang, Yangbangyan and Cao, Xiaochun and Huang, Qingming and Yao, Yuan},
  booktitle={{IEEE} Transactions on Pattern Analysis and Machine Intelligence},
  year={2021}
}