Follow the official instructions to install Detectron (for inference only). This version of Detectron only supports python2 and has tested to support CUDA 8.0 and 9.0 at least. You can skip the steps on Caffe2 if you have torch installed (e.g., in the VLP or GVD conda env) and just finish the rest until here.
For VLP, download the corresponding config file and the checkpoint file and place under this root dir. Refer to extract_feat_flickr30k.sh and tools/extract_features.py for the usage.
For GVD, download the corresponding config file (rename to e2e_faster_rcnn_X-101-64x4d-FPN_2x-gvd.yaml) and the checkpoint file (rename to e2e_faster_rcnn_X-101-64x4d-FPN_2x-gvd.pkl) and place under this root dir. Refer to extract_feat_gvd_anet.sh and tools/extract_features_gvd_anet.py for the usage.