Training end-to-end on my own dataset

I have my own dataset of (image, caption) pairs on which I'd like to train the model. Does this repository enables to do that without first extracting features/bounding boxes?

Can I do it via avoiding passing the flag 
`--features` ?