Consider multiple inputs into model

[MannyKayy]

Currently, the additional tools in chainercv expect that a model takes only one input (an image).
So tools such as apply_prediction_to_iterator.py are only designed to handle single input models. In some cases, a vision model will take multiple inputs (e.g. VQA or Bi-Linear models) and it would be nice to natively support that.

[Hakuyume]

How about adding a new option, n_arg? It indicates the number of additional arguments to passed to predict.

apply_prediction_to_iterator(predict, iterator)  # predict(imgs) is called for each batch
apply_prediction_to_iterator(predict, iterator, n_arg=2)  # predict(imgs, gt0, gt1) is called for each batch

In both cases, it returns imgs, pred_values, gt_values.