How l2-normalization over feature is implemented ?

[ZhuFengdaaa]

this paper states that L2 normalization of the image features is crucial for good performance. However, you just use pool5 data, which is average pooled to become a 2048 vector in generate_tsv.py

I'm wondering if you have implemented L2-normalization over feature or not. If you did, please inform me how you do it. Thanks a lot~

[cengzy14]

I guess the L2 normalization is done in the data loading part of the VQA model part, like https://github.com/asdf0982/vqa-mfb.pytorch/blob/d12291e5219ba29615a130829f520880dd69a211/mfh_coatt_glove/utils/data_provider.py#L254

[ZhuFengdaaa]

Thanks. Yes, it does.