Introduction to EFG Project
EFG, which stands for Efficient, Flexible, and General, is a deep learning framework designed to streamline and innovate research pursuits across various domains within neural networks. With its minimalist design, it allows users to delve into a wide range of research topics using pre-set templates provided by the project. The framework provides ample flexibility and generality, making it a unique tool for researchers.
Recent Developments
EFG has been actively updated with significant enhancements and new releases as follows:
- August 22, 2023: Released code for the ICCV2023 paper: TrajectoryFormer: 3D Object Tracking Transformer with Predictive Trajectory Hypotheses. This resource is available on their GitHub repository.
- April 13, 2023: Introduced support for COCO Panoptic Segmentation using Mask2Former.
- March 30, 2023: EFG is now compatible with PyTorch 2.0.
- March 21, 2023: Released code corresponding to the CVPR2023 Highlight paper titled ConQueR: Query Contrast Voxel-DETR for 3D Object Detection. Additionally, EFG's codebase now supports 2D object detection on the MS COCO dataset and 3D object detection using Waymo and nuScenes datasets.
Installation Guide
Prerequisites
To begin utilizing EFG, the user needs a system with the following:
- GCC 5 (with C++11 or newer)
- Python version 3.6 or higher
- CUDA version 10.1 or higher
- PyTorch version 1.6 or higher
A few specific dependencies are required, like spconv aligned with the user's CUDA version and waymo_open_dataset for Python 3.6 and later versions.
Building from Source
To install EFG from its source code, users should clone the repository, install the necessary packages, and configure the environment for logging of activities like saving model checkpoints and training logs.
Data Preparation
The framework supports data preparation for two primary datasets:
Waymo Dataset
Users need to download specific versions of the Waymo dataset, convert records from tfrecord to pkl format, and set up the dataset structure by creating soft links. These files act as grounds for generating data summaries and the ground truth database from extracted frames.
nuScenes Dataset
Similar to Waymo, there are processes for linking the nuScenes dataset, arranging files correctly, and utilizing specific Python scripts to prepare the data properly.
Getting Started with EFG
Once EFG is set up, users can commence training and evaluating their models. Commands are available for configuring the number of GPUs, resuming tasks, and automatically evaluating models after training.
Model Performance
EFG's framework is tested on high-end NVIDIA A100 GPUs. The documented model performance for both the Waymo Open Dataset in terms of 3D Object Detection and nuScenes is provided, showcasing the results achieved by various integrated methods.
Contributions and Future Work
The EFG project is still burgeoning, and there is ample scope for contributions. Interested individuals can reach out via email to provide inputs or help enhance the project further.
Citation
EFG's work has been recognized, and relevant citations for academic references are documented for various papers affiliated with the framework.
In conclusion, EFG serves as a robust framework fostering further exploration in deep learning while inviting the community to partake in its development.