Cdiscount Image Classification Challange Ongoing Solution.
Right now getting top1 accuray - 0.64 accuracy & top5 accuracy - 0.81
Will provide the weights once challenge is over.
Using an SE-Denset161 model, using intitial iterations with data augmention & last few iterations without any augmentation
use adam_accumulate or sgd_accumulate for batch accumulation, code also provide in the keras issue, keras-team/keras#3556
Learning rate 0-3 = 0.1 4-7 = 0.001 8-10 = 0.0001