This framework can be used to reproduce all experiments performed in “Feature Partitioning for Efficient Multi-Task Architectures”. It provides a variety of functionality for performing and managing deep learning experiments. In particular, it helps manage meta-optimization which is useful when doing hyper-parameter tuning and architecture search.

Please note, this is not an official Google product.

Start a new Cloud Instance from “Deep Learning Image: PyTorch 1.0.0”. Almost everything needed to get the code up and running is included automatically with the Deep Learning Image.

Following instructions assume the git repository has been pulled and placed in the home directory.

Download and set up Visual Decathlon data and annotations:

wget https://www.robots.ox.ac.uk/~vgg/share/decathlon-1.0-devkit.tar.gz
wget https://www.robots.ox.ac.uk/~vgg/share/decathlon-1.0-data.tar.gz
tar zxf decathlon-1.0-devkit.tar.gz
mv decathlon-1.0 ~/mtl/data/decathlon
tar zxf decathlon-1.0-data.tar.gz -C ~/mtl/data/decathlon/data
cd ~/mtl/data/decathlon/data
for f in *.tar; do tar xf "$f"; done

ImageNet data must be set up separately, please check out https://image-net.org/download-images.

Add the following lines to your ~/.bashrc file:

export PATH=~/.local/bin:$PATH
export PYTHONPATH=~/mtl:$PYTHONPATH
ulimit -n 2048

Then run the following:

source ~/.bashrc
pip install --upgrade torch torchvision tensorflow

All code was tested with Python 3.7 and PyTorch 1.0.

A variety of configuration files are available to run different training procedures tested in the paper, some examples include:

python main.py -e single_task_network --config exp.dec --model resnet --task_choice 0
python main.py -e partitioned_mtl_network --config exp.dec --task_choice 1-2-3-4
python main.py -e distillation_test --config exp.dist
python main.py -e es_optimization_test --config exp.es_dist

The argument -e indicates the experiment name, and --config specifies the appropriate configuration file. Further details about network training can be found here.