Skip to content


Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Deep Metric Transfer for Label Propagation with Limited Annotated Data

This repo contains the pytorch implementation for the semi-supervised learning paper (arxiv).


  • Python3: Anaconda is recommended because it already contains a lot of packages:
  • pytorch>=1.0: Refer to
  • other packages: pip install tensorboardX tensorboard easydict scikit-image


  • We formulate semi-supervised learning from a completely different metric transfer perspective.
  • Enjoys the benefit of recent advances self-supervised learning.
  • We hope to draw more attention to unsupervised pretraining for other tasks.

Main results

The test accuracy of our methods and the state-of-the-art methods on CIFAR10 dataset with different number of labeled data.

Method 50 100 250 500 1000 2000 4000 8000
PI-model 27.36 37.20 47.07 56.30 63.70 76.50 84.17 87.30
Mean-Teacher 29.66 36.60 45.49 57.20 65.00 79.00 84.38 87.50
VAT 23.00 35.58 47.61 62.90 72.80 84.00 86.79 88.10
Pseudo-Label 21.00 34.00 45.83 60.30 68.20 78.00 84.79 86.20
Ours 56.34 63.53 71.26 74.77 79.38 82.34 84.52 87.48

Quick start

  • Clone this repo: git clone && cd metric-transfer.pytorch

  • Install pytorch and other packages listed in requirements

  • Download pretrained models and precomputed pseudo labels: bash scripts/ . Make sure the checkpoint folder looks like this:

    |-- pretrain_models
    |   |-- ckpt_instance_cifar10_wrn-28-2_82.12.pth.tar
    |   |-- ... other files
    |   `-- lemniscate_resnet50.pth.tar
    |-- pseudos
    |   |-- instance_nc_wrn-28-2
    |   |   |-- 50.pth.tar
    |   |   |-- ... other files
    |   |   `-- 8000.pth.tar
    |   `-- ... other folders 
    `-- pseudos_imagenet
        `-- instance_imagenet_nc_resnet50
            |-- num_labeled_13000
            |   |-- 10_0.pth.tar
            |   |-- ... other files
            |   `-- 10_9.pth.tar
            `-- ... other folders 
  • Supervised finetune on cifar10 dataset or Imagenet dataset. The cifar dataset will be downloaded automatically. For imagenet, refer to here for details of data preparation.

    # Finetune on cifar
    python \
    	--gpus 0 \
    	--num-labeled 250 \
    	--pseudo-file checkpoint/pseudos/instance_nc_wrn-28-2/250.pth.tar \
    	--resume checkpoint/pretrain_models/ckpt_instance_cifar10_wrn-28-2_82.12.pth.tar \
     	--pseudo-ratio 0.2
    # For imagenet
    n_labeled=13000  # 1% labeled data
    pseudo_ratio=0.1  # use top 10% pseudo label
    python \
        --arch resnet50 \
        --gpus 0,1,2,3 \
        --num-labeled ${n_labeled} \
        --data-dir ${data_dir} \
        --pretrained checkpoint/pretrain_models/lemniscate_resnet50.pth.tar  \
        --pseudo-dir checkpoint/pseudos_imagenet/instance_imagenet_nc_resnet50/num_labeled_${n_labeled} \
        --pseudo-ratio ${pseudo_ratio} \


The proposed method contains three main steps: metric pretraining, label propagation, and supervised finetune.

Metric pretraining

The metric pretraining can be unsupervised or supervised, from the same or different dataset.

We provide code for instance discrimination, which is borrowed from the original pytorch release of instance discrimination. You can run the following command in root director of code to train the instance discrimination on cifar10 dataset:

CUDA_VISIBLE_DEVICES=0 python unsupervised/ \
	--lr-scheduler cosine-with-restart \
	--epochs 1270

For other metric or imagenet dataset, such as colorization on cifar10 dataset, or instance discrimination on imagenet datset, ref to offical released code: colorization, instance discrimination. We also provide the pretrained weight. Refer to scripts/ for more details.

Label propagation

We then can propagation the label using the trained metric from the few labeled examples to a vast collection of unannotated images.

We consider two propagation algorithms: K-nearest neighbors(i.e. knn) and spectral clustering(also called normalized cut, i.e nc). The implementation is in notebooks folder, which is in jupyter notebook format. You can simplely run the notebook to load the weight of metric pretraining approach and propagate to get the pseudo label.

We alse provide the pseudo label for cifar10 and imagenet dataset. Refer to scripts/ for more details.

Supervised finetune

With the estimated pseudo labels on the unlabeled data, we can train a classifier with more data. For simplicity, we omit the confidence weighted supervised training in the current version. Instead, we only use a portion of the most confident pseudo label to training.

Refer to quickstart part for more command instruction.


This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact with any additional questions or comments.


If you find this paper useful in your research, please consider citing:

  title={Deep Metric Transfer for Label Propagation with Limited Annotated Data},
  author={Liu, Bin and Wu, Zhirong and Hu, Han and Lin, Stephen},
  journal={arXiv preprint arXiv:1812.08781},


For any questions, please feel free to create a new issue or reach

Bin Liu:


Deep Metric Transfer for Label Propagation with Limited Annotated Data




Code of conduct





No releases published


No packages published