HierarchicalPCGrad

This repository provides a reimplementation for Gradient Surgery for Multi-Task Learning in PyTorch 1.6.0. The method has further been extended to handle hierarchical task projection and combinations of intra- and inter-priority task projections.

Setup

Install the required packages via:

pip install -r requirements.txt

Usage

import torch
import torch.nn as nn
import torch.optim as optim
from hpcgrad import HPCGrad

# wrap your favorite optimizer
optimizer = HPCGrad(optim.Adam(net.parameters())) 
losses = [...] # a list of per-task losses
priorities = [...] # a list of priorities of tasks, e.g. [1,1,3,2,...]
assert len(losses) == num_tasks
assert len(priorities) == num_tasks
optimizer.pc_backward(losses, priorities) # calculate the gradient can apply gradient modification
optimizer.step()  # apply gradient step

Training

• Mulit-MNIST Please run the training script via the following command. Part of implementation is leveraged from https://github.com/intel-isl/MultiObjectiveOptimization

  python main_multi_mnist.py
  

  The result is shown below.

  Method	left-digit	right-digit
  Jointly Training	90.30	90.01
  PCGrad (this repo.)	95.00	92.00
  PCGrad (official)	96.58	95.50

• Cifar100-MTL coming soon

Reference

Please cite as:

@article{yu2020gradient,
  title={Gradient surgery for multi-task learning},
  author={Yu, Tianhe and Kumar, Saurabh and Gupta, Abhishek and Levine, Sergey and Hausman, Karol and Finn, Chelsea},
  journal={arXiv preprint arXiv:2001.06782},
  year={2020}
}

@misc{Pytorch-PCGrad,
  author = {Wei-Cheng Tseng},
  title = {WeiChengTseng/Pytorch-PCGrad},
  url = {https://github.com/WeiChengTseng/Pytorch-PCGrad.git},
  year = {2020}
}