singhjasdeep/attention-on-attention-for-vqa

Visual Question Answering Project with state of the art single Model performance.

attention-on-attention-for-vqa's Issues

How l2-normalization over feature is implemented ?

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

However, nether did your repository Attention-on-Attention-for-VQA nor the feature exactor repository bottom-up-attention implement the L2-normaliation. I implemented it at the very beginning of the forward procedure by v = v / torch.norm(v, 2). But the validation score decreased by 0.5.

Can anybody explain it ? Thanks~

Access denied to Image Features

Hello,

I tried to download the image features from https://storage.googleapis.com/bottom-up-attention/trainval_36.zip and https://storage.googleapis.com/bottom-up-attention/test2015_36.zip but the server denies access.

<Error>
  <Code>AccessDenied</Code>
  <Message>Access denied.</Message>
  <Details>Anonymous caller does not have storage.objects.get access to bottom-up-attention/test2015_36.zip.</Details>
</Error>

Memory problem while training process

Hello. Launching main.py and recieveing memory error (cannot allocate memory) on the very first epoch. Launching on PC with 64 gb ram, 1080ti and core i9. Does this model requires more than 64 gb? Thanks in advance for your reply.

Can provide the pre-trained model

As titled, not sure you can provide the link for pre-trained model.
Thanks.