This repository contains code for training and evaluating a ResNet or Stacked Hourglass model on the MPII Human Pose dataset using the differentiable spatial to numerical transform (DSNT).
If you want to use DSNT in your own project, check out dsntnn
instead. dsntnn is a small, self-contained library containing all of the operations required
for DSNT, the loss function, and regularization terms.
Edit docker-compose.yml to set the desired location for the MPII Human Pose
dataset on your computer.
Next, download and install the MPII Human Pose dataset:
$ ./run.sh python
>>> from torchdata import mpii
>>> mpii.install_mpii_dataset('/datasets/mpii')
$ ./run.sh pytest- [Optional] Start the Showoff server. Showoff is a visualisation server which can be used to
display metrics while training.
$ docker-compose up -d showoff
- Run the training script (pass
--showoff=""if not using Showoff).$ ./run.sh src/dsnt/bin/train.py --epochs=100
- Wait until the training finishes. If using Showoff, you can monitor progress by going to http://localhost:16676.
bin/infer.py may be used to generate predictions from trained models on the
MPII dataset. The predictions can be written to HDF5 files compatible with
eval-mpii-pose. This is especially
useful for generating Matlab submission files which are compatible with the
official MPII evaluation code.
If you write an implementation of DSNT, please let me know so that I can add it to the list.
- Tensorflow: ashwhall/dsnt
(C) 2017-2018 Aiden Nibali
This project is open source under the terms of the Apache License 2.0.
If you use any part of this work in a research project, please cite the following paper:
@article{nibali2018numerical,
title={Numerical Coordinate Regression with Convolutional Neural Networks},
author={Nibali, Aiden and He, Zhen and Morgan, Stuart and Prendergast, Luke},
journal={arXiv preprint arXiv:1801.07372},
year={2018}
}