============================
Original Paper: "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning"(https://arxiv.org/abs/1602.07261) from Google
Data augmentation applied (please find the data augmentation fork in https://github.com/twtygqyy/caffe-augmentation):
max_color_shift = 5
contrast_variation = 0.8 ~ 1.2
max_brightness_shift = 5
mirror = true
min_side = 328 ~ 480 and crop by 299x299 for training, min_side = 328 and crop by 299x299 for testing
init learning rate = 0.072 with RMSProp optimizer (rms_decay = 0.9 delta = 0.9)
max_iter = 1066080
stepsize = 6663
gamma = 0.94
weight_decay = 0.0004
clip_gradients = 80
4 Geforce 1080 GPU are used for training and batch size = 5 x 4 (Very huge memory required for training)
Test net output #0: accuracy_top1 = 0.729467
Test net output #1: accuracy_top5 = 0.904265
Model link: https://drive.google.com/file/d/0B5i4atpKg9EcOGRqUExXZVNxODQ/view?usp=sharing
Different solver with more iterations is under training right now