/auto_yolo

TensorFlow implementation of Spatially Invariant Attend, Infer, Repeat (SPAIR).

Primary LanguagePythonMIT LicenseMIT

This repository contains code for running experiments from the following paper:

Spatially Invariant Unsupervised Object Detection with Convolutional Neural Networks.
Eric Crawford and Joelle Pineau.
AAAI (2019).

This repository and the companion repository dps are both likely to undergo further development in the future. In general, we will attempt to keep the experiments from the paper runnable, but in case something breaks, one can always check out the aaai_2019 branches of both repositories to obtain the code as it was for the paper. Also, branches for both repos named aaai_2019_v2 preserve most of the behaviour of the original experiments, but with significant code improvements.

Installation

  1. Install TensorFlow with GPU support. auto_yolo should work with any version of TensorFlow > 1.4, but has not been tested extensively with any version other than 1.8.

  2. Clone dps, optionally switch to aaai_2019 or aaai_2019_v2 branches, and install:

    git clone https://github.com/e2crawfo/dps.git
cd dps
(optional: git checkout aaai_2019 or git checkout aaai_2019_v2)
pip install -r requirements.txt
pip install -e .
cd ..

  3. Clone auto_yolo, optionally switch to aaai_2019 or aaai_2019_v2 branches, and install:

    git clone https://github.com/e2crawfo/auto_yolo.git
cd auto_yolo
(optional: git checkout aaai_2019 or git checkout aaai_2019_v2)
pip install -r requirements.txt
pip install -e .
cd ..

  4. Compile custom TensorFlow ops resampler_edge and render_sprites.

    cd auto_yolo/auto_yolo/tf_ops/resampler_edge && make
cd ../render_sprites && make
cd ../../../../

  5. Setup scratch directory and download emnist data.

    cd dps/scripts
python download.py emnist --shape=14,14

    The last line should first ask you for a scratch directory on your computer. It will then download emnist data into that directory, and reshape it to (14, 14) (this process can take a while).

Running Experiments

To train SPAIR on a scattered MNIST dataset:

cd auto_yolo/experiments/comparison
python yolo_air_run.py