Inference + Hugging Face

[mattmdjaga]

Is there an easy way to inference the model on some new examples? Also, are there any plans to put the model on Hugging face?

[viyjy]

Hi, thanks for your interest in our work. What kind of inference do you mean? Image-text retrieval tasks? In terms of putting the model on HuggingFace, I need to check with my team and let you know later. Thanks.

[mattmdjaga]

Inference for the generation task Visual question answering (VQA).

[viyjy]

I see. What't the difficulties of applying our current inference code on general VQA? VQA.py cannot be used in your tasks?

[mattmdjaga]

I tried to extract an inference code from the 'VQA.py' file but I found that you need to supply the model, question and answers. Whereas I thought that the model will generate the answers. So is there no way to inference on VQA without having pre-defined answers?

[viyjy]

The model will generate the answers using an answer decoder, isn't it? The only difference is that we constrain the answer decoder to only generate from the 3,192 candidate answers.