search

airsplay / lxmert

PyTorch code for EMNLP 2019 paper "LXMERT: Learning Cross-Modality Encoder Representations from Transformers".
MIT License
923 stars 157 forks source link

Use of --qaSets parameter #99

Open aurooj opened 3 years ago

aurooj commented 3 years ago

Hi, Thanks for this great code repo.

I have a confusion about use of parameter --qaSets. In the code documentation, it is mentioned param qa_sets: if None, no action o.w., only takes the answers appearing in these dsets and remove all unlabeled data (MSCOCO captions)

Also, we can select qaSets to be vqa, visual7w, gqa Can you please elaborate further? More specifically, I am curious what will happen if I omit one of the options, for instance, passing --qaSets vqa,visual7w will it omit the GQA dataset for VQA pretext task? or my understanding is wrong? Looking forward to hearing from you soon. Thanks in advance!

dreichCSL commented 2 years ago

Did you manage to exclude an entire dataset completely? As far as I understand at this point, specifying --qaSets will only affect which answer classes are trained, but not which datasets are used (i.e. it still uses all datasets, just not all samples, namely not those samples that belong to unsupported answer classes).