efficientnet

EfficientNet for object detection

This repository implements EfficientNet in the SimpleDet framework. Efficient B5 achives the same mAP with ~1/10 FLOPs compared with ResNet-50.

Qucik Start

# train faster r-cnn with efficientnet fpn backbone
python3 detection_train.py --config config/efficientnet/efficientnet_b5_fpn_bn_scratch_400_6x.py

Results and Models

All AP results are reported on minival of the COCO dataset.

Model	InputSize	Backbone	Train Schedule	GPU	Image/GPU	FP16	Train MEM	Train Speed	Box AP	Link
Faster	400x600	B5-FPN	36 epoch(6X)	8X 1080Ti	8	yes	-	75 img/s	37.2	model
Faster	400x600	B5-FPN	54 epoch(9X)	8X 1080Ti	8	yes	-	75 img/s	37.9	-
Faster	400x600	B5-FPN	72 epoch(12X)	8X 1080Ti	8	yes	-	75 img/s	38.3	-

Reference

@inproceedings{tan2019,
  title={EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks},
  author={Tan, Mingxing and Le, Quoc V},
  booktitle={ICML},
  year={2019}
}