Skip to content
Branch: master
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


This repository implements NAS-FPN in the SimpleDet framework.

Qucik Start

# train baseline retinanet following the setting of NAS-FPN
python3 --config config/NASFPN/

# train NAS-FPN
python3 --config config/NASFPN/
python3 --config config/NASFPN/
python3 --config config/NASFPN/

# train hand-crafted neck
python3 --config config/NASFPN/

Results and Models

All AP results are reported on test-dev of the COCO dataset.

Model InputSize Backbone Neck Train Schedule GPU Image/GPU FP16 Train MEM Train Speed Box AP(Mask AP) Link
RetinaNet 640 R50v1b-FPN 1@256 25 epoch 8X 1080Ti 8 yes 6.6G 85 img/s 37.4 model
NAS-FPN 640 R50v1b-FPN 7@256 25 epoch 8X 1080Ti 8 yes 7.8G 66 img/s 40.1 model
NAS-FPN 1024 R50v1b-FPN 7@256 25 epoch 8X 1080Ti 4 yes 9.1G 17 img/s 44.2 model
NAS-FPN 1280 R50v1b-FPN 7@384 25 epoch 8X 1080Ti 2 yes 8.9G 10 img/s 45.3 model
TD-BU* 1280 R50v1b-FPN 3@384 25 epoch 8X 1080Ti 3 yes 10.5G 12 img/s 44.7 model

* Short for TopDown-BottomUp neck which is highly symmetric proposed by Zehao.


  title={NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection},
  author={Ghiasi, Golnaz and Lin, Tsung-Yi and Pang, Ruoming and Le, Quoc V},
You can’t perform that action at this time.