Improved Optimization Strategies for Deep Multi-Task Learning

This code was made under Pytorch v1.2.0.

Each provided folder contains code for its respective dataset (Celeb-A, CityScapes and NYUv2). Inside of these, a single 'train.py' training script is provided for all of the baselines described in the submission paper.

Usage

In the 'train.py' training files, the different arguments can be used to reproduce the benchmark results in the paper. Instructions about their usage can be found in the respective folders.

Data

We use the official releases of each dataset:

• NYUv2: https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html
• Cityscapes: https://www.cityscapes-dataset.com/
• Celeb-A: http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html