Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


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

Multi-Label Learning from Single Positive Labels

Code to reproduce the main results in the paper Multi-Label Learning from Single Positive Labels (CVPR 2021).

Getting Started

See the file in the data directory for instructions on downloading and setting up the datasets.

Training a Model

To train and evaluate a model, run:


Selecting the Training Procedure

To generate different entries of the main table, modify the following parameters:

  1. dataset: Which dataset to use.
  2. loss: Which loss to use.
  3. train_mode: Whether to (a) train a linear classifier on top of pre-extracted features, (b) train end-to-end, or (c) do (a) followed by (b).
  4. val_set_variant: Whether to use a clean val set or a validation set where a single positive is observed for each image.

As written, will run a hyperparameter search over a few different learning rates and batch sizes, save the results for all runs, and report the best run. If desired, modify the code at the bottom of to search over different parameter settings.

The linear_init mode searches over hyperparameters for the fine-tuning phase only. The hyperparameters for the linear training phase are fixed. In particular, linear_init_lr and linear_init_bsize are set to the best learning rate and batch size from a linear_fixed_features hyperparameter search.


  • The requirements.txt files was generated using the wonderful tool pipreqs.
  • Please feel free to get in touch / open an issue if anything is unclear.
  • In this paper we used only those images from NUSWIDE which were still publicly available when we re-crawled the dataset in 2020 using Namhyuk Ahn's downloader. Following the instructions in data/ should yield the exact subset used for our experiments.


If you find our work useful in your research please consider citing our paper:

  title={Multi-Label Learning from Single Positive Labels},
  author={Cole, Elijah and 
          Mac Aodha, Oisin and 
          Lorieul, Titouan and 
          Perona, Pietro and 
          Morris, Dan and 
          Jojic, Nebojsa},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},