Implementation of GoogLeNet by chainer
git clone https://github.com/nutszebra/googlenet.git
cd googlenet
git submodule init
git submodule update
python main.py -p ./ -g 0
-
Data augmentation
Train: Pictures are randomly resized in the range of [256, 512], then 224x224 patches are extracted randomly and are normalized locally. Horizontal flipping is applied with 0.5 probability.
Test: Pictures are resized to 384x384, then they are normalized locally. Single image test is used to calculate total accuracy. -
Auxiliary classifiers
No implementation -
Learning rate
As [[1]][Paper] said, learning rate are multiplied by 0.96 at every 8 epochs. The description about initial learning rate can't be found in [[1]][Paper], so initial learning is setted as 0.0015 that is found in [[2]][Paper2]. -
Weight decay
The description about weight decay can't be found in [[1]][Paper], so by using [[2]][Paper2] and [[3]][Paper3] I guessed that weight decay is 2.0*10^-4.
| network | depth | total accuracy (%) |
|---|---|---|
| my implementation | 22 | 91.33 |
Going Deeper with Convolutions [[1]][Paper]
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift [[2]][Paper2]
Rethinking the Inception Architecture for Computer Vision [[3]][Paper3]
[paper]: https://arxiv.org/abs/1409.4842 "Paper"
[paper2]: https://arxiv.org/abs/1502.03167 "Paper2"
[paper3]: https://arxiv.org/abs/1512.00567 "Paper3"