Skip to content
Unsupervised Feature Learning via Non-parametric Instance Discrimination
Branch: master
Clone or download
Latest commit 34c7ce3 Sep 2, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
datasets init May 4, 2018
models init May 4, 2018
.gitignore init May 4, 2018 readme Jul 27, 2018 cifar resume Jul 27, 2018 imagenet pytorch 0.4 Jul 27, 2018

Unsupervised Feature Learning via Non-parameteric Instance Discrimination

This repo constains the pytorch implementation for the CVPR2018 unsupervised learning paper (arxiv).

  title={Unsupervised Feature Learning via Non-Parametric Instance Discrimination},
  author={Wu, Zhirong and Xiong, Yuanjun and Stella, X Yu and Lin, Dahua},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},


  • We formulate unsupervised learning from a completely different non-parametric perspective.
  • Feature encodings can be as compact as 128 dimension for each image.
  • Enjoys the benefit of advanced architectures and techniques from supervised learning.
  • Runs seamlessly with nearest neighbor classifiers.

Pretrained Model

Currently, we provide pretrained models of ResNet 18 and ResNet 50. Each tar ball contains the feature representation of all ImageNet training images (600 mb) and model weights (100-200mb). You can also get these representations by forwarding the network for the entire ImageNet images.

Nearest Neighbor

Please follow this link for a list of nearest neighbors on ImageNet. Results are visualized from our ResNet50 model, compared with raw image features and supervised features. First column is the query image, followed by 20 retrievals ranked by the similarity.


Our code extends the pytorch implementation of imagenet classification in official pytorch release. Please refer to the official repo for details of data preparation and hardware configurations.

  • supports python27 and pytorch=0.4

  • if you are looking for pytorch 0.3, please switch to tag v0.3

  • clone this repo: git clone

  • Training on ImageNet:

    python DATAPATH --arch resnet18 -j 32 --nce-k 4096 --nce-t 0.07 --lr 0.03 --nce-m 0.5 --low-dim 128 -b 256

    • parameter nce-k controls the number of negative samples. If nce-k sets to 0, the code also supports full softmax learning.
    • nce-t controls temperature of the distribution. 0.07-0.1 works well in practice.
    • nce-m stabilizes the learning process. A value of 0.5 works well in practice.
    • learning rate is initialized to 0.03, a bit smaller than standard supervised learning.
    • the embedding size is controlled by the parameter low-dim.
  • During training, we monitor the supervised validation accuracy by K nearest neighbor with K=1, as it's faster, and gives a good estimation of the feature quality.

  • Testing on ImageNet:

    python DATAPATH --arch resnet18 --resume input_model.pth.tar -e runs testing with default K=200 neighbors.

  • Training on CIFAR10:

    python --nce-k 0 --nce-t 0.1 --lr 0.03


For any questions, please feel free to reach

Zhirong Wu:
You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.