The goal of this challenge is to bring awareness to the energy efficiency of AI accelerators and encourage researchers to innovate a new neural network architecture optimized for AI accelerators.
Participants need to build an efficient machine learning model that completes the given task on the target HW platform as fast as possible.
- Sponsor: Xilinx
- Competition: Online submission through lpcv.ai August 1st 9:00 AM - August 31st 11:59 PM EST
- Restriction: Each team can submit at most once every 24 hours
- Vision Task: Object detection.
- Data: Coco 2017 images data download link: https://cocodataset.org/#download
- Hardware: Ultra96-V2 + Xilinx® Deep Learning Processor Unit (DPU)
- Software: PYNQ
- Reimbursement policy: The first 20 teams that submit solutions will receive $270 per team.
- Technical Support: Please join the online discussion forum lpcvc.slack.com.
- Evaluation: 10^4 / Energy * ReLU (mMAP – 0.2) * ReLU (fps – 5). Where mMAP is INT8 quantized accuracy
- Prizes: Winners will be announced online.
1st place: $1,500.
2nd place: $1,000.
3rd place: $500.
Please register here and send an email to firstname.lastname@example.org to register for the challenge. Please use the subject line
“Subject: [Registration] LPCV 2021 Xilinx Track”
Download latest software:
We list some of the useful references:
- Performance (FPS) and Energy will be measured by running the evaluation.ipynb notebook.
- Accuracy will be measured using codalab online platform. The result JSON file for each team will be uploaded on codalab to get accuracy of the results.
- Please note that we will cross check if the submitted result file was generated from the submitted model file.
- Note: Test dataset is coco2017-testdev
- A valid submission will have the following files in a zipped folder:
- team.json [This is the result file please check format in sample_team.json file, it should be generated using the evaluation.ipynb ]
- lpcv_eval.py [DO NOT MODIFY, please maintain the directory paths for your files as per this script]
- NOTE: To record 5V rail for board power in evaluation.ipynb notebook using pynq.DataRecorder(rails["5V"].power) call, please replace the image.ub on the SD card with this new image.ub file. The PYNQv2.6 image needs some additional config to measure 5V rail. To replace the image.ub file, just plug-in the SD card to a host computer and ‘copy and replace’ the image.ub file.
- Please download the example design files, sample result file, submission notebook and evaluation scripts here.