How to deal with different number of regions feature in one batch?

[njucckevin]

In caption model, we need region features from bottom-up attention model. But if we got different shape of features, for example, extract 10 regions from image A and 20 regions from image B, how can I train them in ONE batch?

[MeiqiGuo]

I think the feature files suggested by this repo have fix number of features per image -- 36.