Event-based Asynchronous Sparse Convolutional Networks

This is the code for the paper Event-based Asynchronous Sparse Convolutional Networks (PDF) by Nico Messikommer*, Daniel Gehrig*, Antonio Loquercio, and Davide Scaramuzza.

If you use any of this code, please cite the following publication:

@InProceedings{Messikommer20eccv,
  author        = {Nico Messikommer and Daniel Gehrig and Antonio Loquercio and Davide Scaramuzza},
  title         = {Event-based Asynchronous Sparse Convolutional Networks},
  journal       = {European Conference on Computer Vision. (ECCV)},
  url           = {http://rpg.ifi.uzh.ch/docs/ECCV20_Messikommer.pdf},
  year          = 2020
}

Installation

First set up an Anaconda environment:

conda create -n asynet python=3.7  
conda activate asynet

Then clone the repository and install the dependencies with pip

git clone git@github.com:uzh-rpg/rpg_asynet.git
cd rpg_asynet/
pip install -r requirements.txt

In addition sparseconvnet 0.2 needs to be installed from here.

CPP Bindings

To build the cpp bindings for the event representation tool, you can follow the instructions below:

pip install event_representation_tool/

For the bindings for asynchronous sparse convolutions, we first need to clone the 3.4.0-rc version of Eigen into the include folder. In addition, pybind11 is required.

cd async_sparse_py/include
git clone https://gitlab.com/libeigen/eigen.git --branch 3.4.0-rc1
conda install -c conda-forge pybind11

Finally, the bindings can be installed

pip install async_sparse_py/

Sparse CNN Training

The training parameters can be adjusted in the config/settings.yaml file. The following training tasks and datasets are supported:

Classification on NCars and NCaltech101.
Object Detection on Prophesee Gen1 Automotive and NCaltech101

Download location of the datasets:

To test the code, make a directory data/ in the root of the repository and download one of the datasets:

mkdir data
cd data
wget http://rpg.ifi.uzh.ch/datasets/gehrig_et_al_iccv19/N-Caltech101.zip
unzip N-Caltech101.zip
rm N-Caltech101.zip

The dataset can be configured in the dataset/name tag in config/settings.yaml. Different model types can be chosen based on the task and whether or not sparse convolutions should be used.

Sparse VGG for Classification and Object Detection
Standard VGG for Classification and Object Detection

These can be configured in the model tag in config/settings.yaml.

The following command starts the training:

CUDA_VISIBLE_DEVICES=<GPU_ID>, python train.py --settings_file config/settings.yaml

By default, a folder with the current date and time is created in log/ containing the corresponding tensorboard files.

Unit Tests

To test the different asynchronous and sparse layers, multiple unit tests are implemented in unittests/:

Asynchronous Sparse Convolution Layer 2D: sparse_conv2D_test.py
Asynchronous Sparse Convolution Layer 2D CPP Implementation: sparse_conv2D_cpp_test.py
Asynchronous Sparse Max Pooling Layer: sparse_max_pooling_test.py
Asynchronous Sparse VGG: sparse_VGG_test.py.
There are three paths in the script sparse_VGG_test.py specified with 'PATH_TO_MODEL' and 'PATH_TO_DATA', which need to be replaced with the NCaltech Classification dataset and the sparse classification model trained on N-Caltech.

To run the unittests call

python -m unittest discover -s unittests/ -p '*_test.py'

Evaluation

There is one script in the evaluation/sliding_window_flops.py folder for computing the number of FLOPs. The command to execute the script is:

python -m evaluation.sliding_window_flops --setting config/settings.yaml
    --save_dir <PATH_TO_DIR> --num_events 1 --num_samples 500
    --representation histogram  --use_multiprocessing

The output of the script are the numbers of FLOPs for the four processing modes (Asyn Sparse Conv, Asyn Conv, Sparse Conv, Standard Conv).

rolsheng/rpg_asynet