search

tohinz / semantic-object-accuracy-for-generative-text-to-image-synthesis

Code for "Semantic Object Accuracy for Generative Text-to-Image Synthesis" (TPAMI 2020)
MIT License
106 stars 23 forks source link

What conference is this paper submitted? #9

Closed patrickphat closed 4 years ago

patrickphat commented 4 years ago

Hi Tohinz and your team :D

I'm currently an undergrad and very interested in Text-to-Image synthesis. I find your paper very fascinating:

The idea really makes sense to me, so I want to use this paper in my thesis, but only if this paper is published in a conference (my university's rule). Would you mind sharing me the expected date of this paper to be submitted to the conference?

Also, if I found some bugs in the code, I can join as a contributor :P Thank you!!

tohinz commented 4 years ago

Hi, the paper is currently under review with a journal. We're in the second round already and have addressed the reviewer's feedback from the first round. I hope to hear back from the journal soon (the reviews are done and the paper is currently with the associate editor). I will let you know once I have updates.

Sure, if you find bugs just do a pull request.

tohinz commented 4 years ago

Also, if you don't have time to wait until we get feedback from the journal (and I obviously don't know for sure that it will be accepted), you can also look into our previous paper published at ICLR 2019. The architecture itself is similar and the code is available here: https://github.com/tohinz/multiple-objects-gan

patrickphat commented 4 years ago

Hi @tohinz , I'm now looking at the published paper at ICLR 2019 and also your blog!! Also, how long does it take to train an OP-GAN?

tohinz commented 4 years ago

Hi, training the OP-GAN on COCO with the current parameters takes quite long (about two weeks). However, I believe you can reduce the number of parameters without much (or any) loss in the final quality. I didn't test this, but I think you could easily use half has many filters in the global and OP pathways and still get the same results. If you use fewer filters you might also be able to increase the batch size (you might have to adapt the learning rate then) to speed up the training even more.

patrickphat commented 4 years ago

I'll try implementing your suggestion, thank you @tohinz

tohinz commented 4 years ago

FYI: we just got the acceptance note from TPAMI for this paper, though it might still be a couple of weeks until it is officially available from their website.

tohinz commented 4 years ago

Here is the preprint + DOI on the TPAMI website: https://ieeexplore.ieee.org/document/9184960