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


Pretrained masks and backbones are available here:

Datasets in PyTorch format are available here:
All rights belong to the respective publishers. The datasets are provided only to aid reproducibility.

The PyTorch-friendly Places365 dataset can be downloaded from

Place masks in checkpoints/ and unzipped datasets in data/

VGG-16 ResNet-50 DenseNet-121
CUBS 20.75 18.23 19.24
Stanford Cars 11.78 10.19 10.62
Flowers 6.93 4.77 4.91
WikiArt 29.80 28.57 29.33
Sketch 22.30 19.75 20.05

Note that the numbers in the paper are averaged over multiple runs for each ordering of datasets. These numbers were obtained by evaluating the models on a Titan X (Pascal). Note that numbers on other GPUs might be slightly different (~0.1%) owing to cudnn algorithm selection.


Python 2.7 or 3.xx
torchnet (pip install git+
tqdm (pip install tqdm)

Run all code from the src/ directory, e.g. ./scripts/


Check out src/scripts/

This script uses the default hyperparams and trains a model as described in the paper. The best performing model on the val set is saved to disk. This saved model includes the real-valued mask weights.

By default, we use the models provided by torchvision as our backbone networks. If you intend to evaluate with the masks provided by us, please use the correct version of torch and torchvision. In case you want to use a different version, but still want to use our masks, then download the pytorch_backbone networks provided in the box link above. Make appropriate changes to your pytorch code to load those backbone models.

Saving trained masks only.

Check out src/scripts/

This extracts the binary/ternary masks from the above trained models, and saves them separately.


Use the saved masks, apply them to a backbone network and run eval.

By default, our backbone models are those provided with torchvision.
Note that to replicate our results, you have to use the package versions specified above.
Newer package versions might have different weights for the backbones, and the provided masks won't work.

cd src  # Run everything from src/

CUDA_VISIBLE_DEVICES=0 python --mode eval --dataset flowers \
  --arch vgg16 \
  --maskloc ../checkpoints/


Code for Piggyback: Adapting a Single Network to Multiple Tasks by Learning to Mask Weights






No releases published


No packages published