What's this

Implementation of GoogLeNet-v3 [1] by chainer

Dependencies

git clone https://github.com/nutszebra/googlenet_v3.git
cd googlenet_v3
git submodule init
git submodule update

How to run

python main.py -p ./ -g 0 

Details about my implementation

• 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

• Gradient clipping
  2.0

• SGD momentum lr=0.1, momentum=0.9

• Learning rate
  Initial learning rate is 0.1 acoording to [1], and it is multiplied by 0.94 at every 2 epochs.

• Weight decay
  According to [2], weight decay is 4.0*10^-5.

Cifar10 result

network	depth	total accuracy (%)
my implementation	49	94.74

References

Rethinking the Inception Architecture for Computer Vision [1]
Xception: Deep Learning with Depthwise Separable Convolutions [2]