Pytorch Contrastive and Triplet Loss experiments

Setup

conda install --file requirements.txt

Run experiments

python main.py

Results (mAP@100)

Dataset	Contrastive Loss	Triplet Loss	Batch Hard
MNIST	0.986	0.983	--
FashionMNIST	0.86	0.871	--
CIFAR10	0.697	0.639	--
Cars3D	0.501	0.532	0.667
CarsEPFL	0.832	0.769	0.761
CarsShapeNet	0.56	0.679	0.739

Loss Implementations

1. Contrastive Loss
2. Vanilla Triplet loss
3. Batch Hard Triplet Loss
4. Batch Soft Triplet loss

DataLoaders

1. MNIST
2. FashionMNIST
3. CIFAR10
4. Cars3D
5. CarsEPFL
6. CarsShapeNet

References

1. Github Adambielski's siamese-triplet
2. Github Beyond-Binary-Supervision-CVPR19
3. Github kilsenp's triplet-reid-pytorch
4. Data Cars3D
5. Data CarsEPFL
6. Data CarsShapeNet
7. Paper FaceNet
8. Paper In Defense of Triplet Loss

TODO

1. Argparser