This repository is the Pytorch implementation of "A Comprehensive Approach to Unsupervised Embedding Learning based on AND Algorithm"
- Illustration of basic architecture in Super-AND.
- We efficiently unify various techniques from state-of-the-art models and introduce a new loss, AE-loss, considers the adversarial training on the unsupervised embedding learning task.
- Super-AND outperforms all baselines in various benchmark datasets. It achieved an accuracy of 89.2% in the CIFAR-10 dataset with the Resnet18 backbone network, compared to the state-of-the-art that gained 86.3%.
- The extensive experiments and ablation study show that every component in Super-AND contributes to the performance increase and also indicate their synergies are critical.
T 60EF he proposed Super-AND outperforms state-of-the-art baselines and following tables are the main results evaluated on two kinds of datasets: coarse-grained and fine-grained.
- Table 1: kNN Evaluation on coarse-grained datasets
- Table 2: kNN Evaluation on fine-grained datasets
Qualitative analysis is conducted to verify our model's superiority. We compared the original AND algorithm and our proposed model Super-AND.
- Figure 1: The nearest retrievals of example queries from STL-10. The upper retrieval row from every query shows the results from the AND model, and the lower ones are from the Super-AND model. The left-side results are successful cases for both models, and the right-side results are failure cases. Images with surrounding red frames indicate the wrongly retrieved negative samples.
- Figure 2: Class probabilities predicted from both AND and Super-AND model on example STL-10 images are illustrated as bars. The green bar represents correctly predicted answers, and the yellow bar represents incorrectly predicted answers.
We support the pretrained model for ResNet18 on CIFAR10 dataset that reached the state-of-the-art accuracy in our paper "Super-AND: A Holistic Approach to Unsupervised Embedding Learning".
- Clone the repository
git clone https://github.com/super-AND/super-AND.git
- Training on CIFAR10
python3 main.py --dataset cifar10 --network resnet18
-
(option) Reproduce the result in paper
Download the file mentioned in pretrained model part then put it to checkpoint dir.
python3 main.py --resume ckpt_SAND.t7 --test_only