How to speed up the data load when training the ImageNet on Keras?

I just making a parallel load program for ImageNet Data load (https://github.com/littlelove2013/ImageNetTrainingOnKeras.git), But the GPU usage is still low! I try to find the reason and training the AlexNet just using one batch in the memory. I found that, the usage of GPU is still low and even to 0% at each 3s. The most high usage is 50%. How to improve the GPU usage when trainin with Keras?