An Unoffical Implementation of PeleeNet by TensorFlow, Keras
An Unoffical Implementation of PeleeNet by TensorFlow, Keras.
Implemented training with CIFAR-10.
R. J. Wang, X. Li, C. X. Ling. Pelee: A Real-Time Object Detection System on Mobile Devices. NIPS. 2018.
https://arxiv.org/abs/1804.06882
The original implementation by Caffe.
https://github.com/Robert-JunWang/Pelee
from pelee_net import PeleeNet
model = PeleeNet(input_shape=(224,224,3), use_stem_block=True, n_classes=1000)
Augmentation | Stem Block | No weight Decay | 5e-4 Weight Decay |
---|---|---|---|
No | No | 0.7633 | 0.9247 |
No | Yes | 0.8280 | 0.9218 |
Yes | No | 0.8881 | 0.9446 |
Yes | Yes | 0.8996 | 0.9410 |
Data augmentation is the standard data augmentation(4 pixels shift and horizontal flip).
There was no discussion int the paper on weight decay. But I noticed that weight decay is important in increasing the accuracy of CIFAR-10, so I added it.
DenseNetの軽量版、PeleeNetをKerasで実装した
https://qiita.com/koshian2/items/187e240f478504079e7a