Awesome
Multi-level Scene Description Network
This is our implementation of Multi-level Scene Description Network in Scene Graph Generation from Objects, Phrases and Region Captions. The project is based on PyTorch version of faster R-CNN. (Update: model links have been updated. Sorry for the inconvenience.)
*Updates*
We have released our newly proposed scene graph generation model in our ECCV-2018 paper:
Factorizable Net: An Efficient Subgraph-based Framework for Scene Graph Generation.
Check the github repo Factorizable Net if you are interested.
Progress
- README for training
- README for project settings
- our trained RPN
- our trained Full Model
- Our cleansed Visual Genome Dataset
- training codes
- evaluation codes
- Model acceleration (please refer to our ECCV project).
- Multi-GPU support: we have release a beta Multi-GPU version of our FactorizableNet. If you want to enhance the training speed, please check that project.
We are still working on the project. If you are interested, please Follow our project.
Project Settings
-
Install the requirements (you can use pip or Anaconda):
conda install pip pyyaml sympy h5py cython numpy scipy conda install -c menpo opencv3 conda install -c soumith pytorch torchvision cuda80 pip install easydict
-
Clone the Faster R-CNN repository
git clone git@github.com:yikang-li/MSDN.git
-
Build the Cython modules for nms and the roi_pooling layer
cd MSDN/faster_rcnn ./make.sh cd ..
-
Download the trained full model and trained RPN, and place it to
output/trained_model
-
Download our cleansed Visual Genome dataset. And unzip it:
tar xzvf top_150_50.tgz
- p.s. Our ipython scripts for data cleansing is also released.
-
Download Visual Genome images
-
Place Images and cleansed annotations to coresponding folders:
mkdir -p data/visual_genome
cd data/visual_genome
ln -s /path/to/VG_100K_images_folder VG_100K_images
ln -s /path/to/downloaded_folder top_150_50
- p.s. You can change the default data directory by modifying
__C.IMG_DATA_DIR
infaster_rcnn/fast_rcnn/config.py
Training
-
Training in multiple stages. (Single-GPU training may take about one week.)
- Training RPN for object proposals and caption region proposals (the shared conv layers are fixed). We also provide our pretrained RPN model.
by default, the training is done on a small part of the full dataset:
CUDA_VISIBLE_DEVICES=0 python train_rpn.py
For full Dataset Training:
CUDA_VISIBLE_DEVICES=0 python train_rpn.py --max_epoch=10 --step_size=2 --dataset_option=normal --model_name=RPN_full_region
--step_size
is set to indicate the number of epochs to decay the learning rate,dataset_option
is to indicate the\[ small | fat | normal \]
subset.- Training MSDN
Here, we use SGD (controled by
--optimizer
)by default:CUDA_VISIBLE_DEVICES=0 python train_hdn.py --load_RPN --saved_model_path=./output/RPN/RPN_region_full_best.h5 --dataset_option=normal --enable_clip_gradient --step_size=2 --MPS_iter=1 --caption_use_bias --caption_use_dropout --rnn_type LSTM_normal
-
Furthermore, we can directly use end-to-end training from scratch (not recommended). The result is not good.
CUDA_VISIBLE_DEVICES=0 python train_hdn.py --dataset_option=normal --enable_clip_gradient --step_size=3 --MPS_iter=1 --caption_use_bias --caption_use_dropout --max_epoch=11 --optimizer=1 --lr=0.001
Evaluation
Our pretrained full Model is provided for your evaluation for further implementation. (Please download the related files in advance.)
./eval.sh
Currently, the accuracy of our released version is slightly different from the reported results in the paper:Recall@50: 11.705%; Recall@100: 14.085%.
Acknowledgement
We thank longcw for his generously releasing the PyTorch Implementation of Faster R-CNN.
Reference
@inproceedings{li2017msdn,
author={Li, Yikang and Ouyang, Wanli and Zhou, Bolei and Wang, Kun and Wang, Xiaogang},
title={Scene graph generation from objects, phrases and region captions},
booktitle = {Proceedings of the IEEE International Conference on Computer Vision},
year = {2017}
}
License:
The pre-trained models and the MSDN technique are released for uncommercial use.
Contact Yikang LI if you have questions.