Skip to content
Switch branches/tags


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

AdaShare: Learning What To Share For Efficient Deep Multi-Task Learning (NeurIPS 2020)


alt text

AdaShare is a novel and differentiable approach for efficient multi-task learning that learns the feature sharing pattern to achieve the best recognition accuracy, while restricting the memory footprint as much as possible. Our main idea is to learn the sharing pattern through a task-specific policy that selectively chooses which layers to execute for a given task in the multi-task network. In other words, we aim to obtain a single network for multi-task learning that supports separate execution paths for different tasks.

Here is the link for our arxiv version.

Welcome to cite our work if you find it is helpful to your research.

  title={Adashare: Learning what to share for efficient deep multi-task learning},
  author={Sun, Ximeng and Panda, Rameswar and Feris, Rogerio and Saenko, Kate},
  journal={Advances in Neural Information Processing Systems},

Experiment Environment

Our implementation is in Pytorch. We train and test our model on 1 Tesla V100 GPU for NYU v2 2-task, CityScapes 2-task and use 2 Tesla V100 GPUs for NYU v2 3-task and Tiny-Taskonomy 5-task.

We use python3.6 and please refer to this link to create a python3.6 conda environment.

Install the listed packages in the virual environment:

conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
conda install matplotlib
conda install -c menpo opencv
conda install pillow
conda install -c conda-forge tqdm
conda install -c anaconda pyyaml
conda install scikit-learn
conda install -c anaconda scipy
pip install tensorboardX


Please download the formatted datasets for NYU v2 here

The formatted CityScapes can be found here.

Download Tiny-Taskonomy as instructed by its GitHub.

The formatted DomainNet can be found here.

Remember to change the dataroot to your local dataset path in all yaml files in the ./yamls/.


Policy Learning Phase

Please execute for policy learning, using the command

python --config <yaml_file_name> --gpus <gpu ids>

For example, python --config yamls/adashare/nyu_v2_2task.yml --gpus 0.

Sample yaml files are under yamls/adashare

Note: use domainnet branch for experiments on DomainNet, i.e. python --config <yaml_file_name> --gpus <gpu ids>

Retrain Phase

After Policy Learning Phase, we sample 8 different architectures and execute for retraining.

python --config <yaml_file_name> --gpus <gpu ids> --exp_ids <random seed id>

where we use different --exp_ids to specify different random seeds and generate different architectures. The best performance of all 8 runs is reported in the paper.

For example, python --config yamls/adashare/nyu_v2_2task.yml --gpus 0 --exp_ids 0.

Note: use domainnet branch for experiments on DomainNet, i.e. python --config <yaml_file_name> --gpus <gpu ids>


After Retraining Phase, execute for get the quantitative results on the test set.

python --config <yaml_file_name> --gpus <gpu ids> --exp_ids <random seed id>

For example, python --config yamls/adashare/nyu_v2_2task.yml --gpus 0 --exp_ids 0.

We provide our trained checkpoints as follows:

  1. Please download our model in NYU v2 2-Task Learning
  2. Please donwload our model in CityScapes 2-Task Learning
  3. Please download our model in NYU v2 3-Task Learning

To use these provided checkpoints, please download them to ../experiments/checkpoints/ and uncompress there. Use the following command to test

python --config yamls/adashare/nyu_v2_2task_test.yml --gpus 0 --exp_ids 0
python --config yamls/adashare/cityscapes_2task_test.yml --gpus 0 --exp_ids 0
python --config yamls/adashare/nyu_v2_3task_test.yml --gpus 0 --exp_ids 0

Test with our pre-trained checkpoints

We also provide some sample images to easily test our model for nyu v2 3 tasks.

Please download our model in NYU v2 3-Task Learning

Execute to test on sample images in ./nyu_v2_samples, using the command

python --config  yamls/adashare/nyu_v2_3task_test.yml --gpus 0

It will print the average quantitative results of sample images.


If any link is invalid or any question, please email


AdaShare: Learning What To Share For Efficient Deep Multi-Task Learning



No releases published


No packages published