You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is it possible to find the HDF5 file somewhere? I would like to run and test the algorithm trained on the Visual Genome dataset that you also used in your paper. I read that the processing of such a dataset will produce a HDF5 file that is larger than 100GB. is this the real model that then must be trained or it is a kind of special representation of the data?
Moreover I would like to know how different is the pre-trained model that you provided here on github: how many images and region descriptions are you using here? Do you think that using the real Visual Genome dataset would make any significant differences?
The text was updated successfully, but these errors were encountered:
Is it possible to find the HDF5 file somewhere? I would like to run and test the algorithm trained on the Visual Genome dataset that you also used in your paper. I read that the processing of such a dataset will produce a HDF5 file that is larger than 100GB. is this the real model that then must be trained or it is a kind of special representation of the data?
Moreover I would like to know how different is the pre-trained model that you provided here on github: how many images and region descriptions are you using here? Do you think that using the real Visual Genome dataset would make any significant differences?
The text was updated successfully, but these errors were encountered: