search

wenhuchen / Meta-Module-Network

Code for WACV 2021 Paper "Meta Module Network for Compositional Visual Reasoning"
43 stars 6 forks source link

Code for extracting features from Bottom up attention #8

Closed sumit-agrwl closed 3 years ago

sumit-agrwl commented 3 years ago

Can you please provide the code for extracting features using bottom up code? I am working on a small subset of images and want to replicate the numpy files that you have used as features?

wenhuchen commented 3 years ago

@linjieli222

linjieli222 commented 3 years ago

@sumit-agrwl We follow the same script used in UNITER to extract the image features. Check out the step 6 in Quick Start. Note that the docker for the script extract_imgfeat.sh is built with caffe that is compatible only with older cuda drivers.

linjieli222 commented 3 years ago

If it is hard to downgrade your cuda driver to be compatible with the docker, there are a lot of alternative pytorch bottom-up attention repos, such as https://github.com/MILVLG/bottom-up-attention.pytorch, which might not give you the features that is exactly the same as what we have provided, but the quality of these features should be comparable or even better.

linjieli222 commented 3 years ago

More details about the docker are explained in this issue: https://github.com/ChenRocks/UNITER/issues/18