csalt-research/perturb-predict-paraphrase

Implementation of Perturb, Predict & Paraphrase: Semi-supervised Learning using Noisy Student for Image Captioning

perturb-predict-paraphrase's Issues

questions about new_unlab.json and the unlabel data missing

Thank you very much for opening such great work.

When I run the trainstudent.py, I found that the unlabel data was missing. I want to know how new_unlab.json is obtained and how can I get the data in new_unlab.json. I has download the cocobu_att, cocobu_fc, cocotalk_att, cocotalk_fc.

No such file or directory: '../172679.npz' This is the error and I can not find the data of new_unlab.json in cocobu and cocotalk.

Where does unpaired captions come from?

Hi! It's pleasure to see your paper and code. But I'm confused when I see it in section 5.5 Paraphrasing Part.

The paraphraser is trained on the set of unpaired captions UC

So, the question is: What is the unpaired captions used in your experiments?
As we know that Unlabeled MSCOCO splits only have images without captions.
Looking forward for your reply! Thanks again!

OSError: beampreds.txt not found.

Thank you very much for opening such great work.

When I run the get_pseudo_labels.py, I found that the 'beampreds.txt' cited by BuDataLoader is missing and caused relevant OSError. I want to know how beampreds.txt is obtained and how can I generate the labelfile with my dataset.