I use Keras VGG16 with the imagenet fine tuned model to classify natural scene images to 5 classes, but now my validation accuracy oscillating at about 85%, cannot be even better unless keep on training to overfit, how can i do to improve the performance, are there any experienced friend can provide some helpful suggestions? The target is up to 90% validation accuracy or more …
I thought some tricks, like lr decay, class weight balance,
My training set contains 12000 pics, and validation contains 3000 pics. I used Adam with lr = 1e-7.
Really appreciate…