Skip to content

Latest commit

 

History

History
63 lines (54 loc) · 2.98 KB

README.md

File metadata and controls

63 lines (54 loc) · 2.98 KB
Raw

AEGNN: Asynchronous Event-based Graph Neural Networks

AEGNN

This repository contains code from our 2022 CVPR paper AEGNN: Asynchronous Event-based Graph Neural Networks by Simon Schaefer*, Daniel Gehrig*, and Davide Scaramuzza. If you use our code or refer to this project, please cite it using

@inproceedings{Schaefer22cvpr,
  author    = {Schaefer, Simon and Gehrig, Daniel and Scaramuzza, Davide},
  title     = {AEGNN: Asynchronous Event-based Graph Neural Networks},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
  year      = {2022}
}

Installation

The code heavily depends on PyTorch and the PyG framework, which is optimized only for GPUs supporting CUDA. For our implementation the CUDA version 11.3 is used. Install the project requirements with:

conda env create --file=environment.yml

Processing Pipeline

We evaluated our approach on three datasets. NCars, NCaltech101 and Prophesee Gen1 Automotive. Download them and extract them. By default, they are assumed to be in /data/storage/, this can be changed by setting the AEGNN_DATA_DIR environment variable.

Pre-Processing

To efficiently train the graph neural networks, the event graph is generated offline during pre-processing. For specific instructions about the data structure and data pre-processing, please refer to the dataset's readme.

Asynchronous & Sparse Pipeline

The code allows to make any graph-based convolutional model asynchronous & sparse, with a simple command and without the need to change the model's definition or forward function.

>>> import aegnn
>>> model = GraphConvModel()
>>> model = aegnn.asyncronous.make_model_asynchronous(model, **kwargs)

We support all graph convolutional layers, max pooling, linear layers and more. As each layer is independently transformed to work asynchronously and sparsely, if there is a layer, that we do not support, its dense equivalent is used instead.

Evaluation

We support automatic flops and runtime analysis, by using hooking each layer's forward pass. Similar to the make_model_asynchronous() function, among other, all graph-based convolutional layers, the linear layer and batch normalization are supported. As an example, to run an analysis of our model on the NCars dataset, you can use:

python3 aegnn/evaluation/flops.py --device X

Contributing

If you spot any bugs or if you are planning to contribute back bug-fixes, please open an issue and discuss the feature with us.