mknbv / meta

Code for COLT'22 paper "Trace norm regularization for multi-task learning with scarce data"

Code for paper "Trace norm regularization for multi-task learning with scarce data"

To install the package:

pip install -e .

Running multiple experiments simultaneously is done through task spooler. To install it run:

sudo apt-get install task-spooler

Running meta script generates experiment files from which the plots could be produced as done in the notebooks/plots.ipynb Jupyter notebook. To run figure 1 experiments:

meta --ntasks-range 100 6400 7 --task-size 10 \
    --num-runs 12 --logdir logdir/figure.01

figure 2 experiments:

meta --ntasks-range 100 6400 7 --task-size 25 \
    --num-runs 12 --logdir logdir/figure.02

figure 3 experiments:

meta --task-size-range 5 30 9 --ntasks 800 \
    --num-runs 12 --logdir logdir/figure.03

figure 4 experiments:

meta --ntasks-range 100 6400 7 --features-dist adversarial \
    --task-size 25 --num-runs 12 --logdir logdir/figure.04

figure 5 experiments:

meta --label-scale-range 1e-3 2 14 --ntasks 800 --task-size 10 \
    --num-runs 12 --logdir logdir/figure.05