SGG_Ex_RC
SGG_Ex_RC copied to clipboard
Published
20 hours ago •
arxrean
arxrean
Code for Scene Graph Generation with External Knowledge and Image Reconstruction
Notice
-
This is an unofficial implementation for CVPR19:Scene Graph Generation with External Knowledge and Image Reconstruction.
-
Code is based on FactorizableNet.
-
All modules in the paper have been done. Code can run without error. Only support VRD dataset now.
-
There is still something wrong with training loss for original code in FactorizableNe. If I train the model with only RPN pretrained, all proposals will be predicted as 0(background) after one iteration, with or without added modules in this repo.
Instruction
- Download VRD data as same as in FactorizableNet.
- Compile the Faster-RCNN lib as same as in FactorizableNet.
- Download Glove txt from GloVe and put it into ./data/Glove. glove.840B.300d.zip is used.
- CUDA_VISIBLE_DEVICES=0 python train_FN.py --path_opt options/models/VRD.yaml --rpn output/RPN_VRD.h5 --CON_use 0 | 1 (Concept network for getting KB and refining object proposals.) --RC_use 0 | 1 (Reconstruct images from object proposals) --GAN_use 0 | 1 (G/D training)
arxrean