PRObablistic Parametric rEgression Loss (PROPEL)

PRObabilistic Parametric rEgresison Loss (PROPEL) is a loss function that enables probabilisitic regression for a neural network. It achieves this by enabling a neural network to learn parameters of a mixture of Gaussian distribution.

Further details about the loss can be found in the paper: PROPEL: Probabilistic Parametric Regression Loss for Convolutional Neural Networks

This repository provides official pytorch implementation of PROPEL.

Installation Instructions

PROPEL can be installed using the following command

pip install torchpropel

pip install git+https://github.com/masadcv/PROPEL.git

Usage Example

import torch
import numpy as np

from torchpropel import PROPEL

# Our example has a neural network with
# output [num_batch, num_gaussians, num_dims]
num_batch = 4
num_gaussians = 6
num_dims = 3

# setting ground-truth variance sigma_gt=0.2
sigma_gt = 0.2
propel_loss = PROPEL(sigma_gt)

# ground truth targets for loss
y = torch.ones((num_batch, num_dims)) * 0.5

# example prediction - this can also be coming as output of a neural network
feat_g = np.random.randn(num_batch, num_gaussians, 2 * num_dims) * 0.5
feat_g[:, :, num_dims::] = 0.2
feat = torch.tensor(feat_g, dtype=y.dtype)

# compute the loss
L = propel_loss(feat, y)

print(L)

Documentation

Further details of each function implemented for PROPEL can be accessed at the documentation hosted at: https://masadcv.github.io/PROPEL/index.html.

Citing PROPEL

Pre-print of PROPEL can be found at: PROPEL: Probabilistic Parametric Regression Loss for Convolutional Neural Networks

If you use PROPEL in your research, then please cite:

BibTeX:

@inproceedings{asad2020propel,
  title={PROPEL: Probabilistic Parametric Regression Loss for Convolutional Neural Networks},
  author={Asad, Muhammad and Basaru, Rilwan and Arif, SM and Slabaugh, Greg},
  booktitle={25th International Conference on Pattern Recognition},
  pages={},
  year={2020}}