vegb/inlg

Implementation of "Visualize Before You Write: Imagination-Guided Open-Ended Text Generation".

inlg's Issues

A bug in generation

Hi, I found the code may be wrong.

iNLG/code/trainer.py

Line 134 in 96800ea

"max_length": self._max_length if self._max_length is not None else self.model.config.max_length,

First, if self does not have an attribute _max_length, directly calling self._max_length leads to an error.
So, I changed it to hasattr(self, '_max_length').

After that I found this condition always goes to the else option, so the max_length will be model.config.max_length, which equals 20 by default. I think it should args.max_output_length, or it may cause a truncation when the task is story generation.

How to get image features of my own dataset

Thanks for contributing such interesting and meaningful work!

Your codes work great, and I wonder how I can use the code for my own dataset. I noticed you have processed vitb32.hdf5 for each dataset, which contains a 512-dimensional ViT feature for each image. So could you give some official instructions to create a vitb32.hdf5 for my own dataset?