HDF5 with Python's h5py

Question

I am tackling the big data image recognition and will be training a neural net to recognize car brand logotypes. I have scraped some ~30k images of 40 brands and am about to decide on the data storage format. In order for the net not to overfit to such a small data, I will do some augmentation, but this will increase the big file handling problem...

I heard that HDF5 is able to efficiently do it, but has anyone of you got aome experience with doing this in Python? Is it optimal for this amount of data?