search

Cadene / vqa.pytorch

Visual Question Answering in Pytorch
716 stars 177 forks source link

Hi, Cadene, why don't you update the feature extractor when training? #36

Open zengxianyu opened 6 years ago

zengxianyu commented 6 years ago

Hi, Cadene,

Thank you very much for your code, it's very helpful. But it seems that you use resnet as a seperate feature extractor instead of updating it when training the VQA model. Could I ask you the reason of that? Intuitively, it might give better results to train the feature extractor and the VQA model together, since the size of training data is not small.

Thank you in advance.

Cadene commented 6 years ago

@zengxianyu very good question.

Actually, we tried and it did not work well (marginal gain for an expensive increase in training time).