gsig / srnn Goto Github PK

Update 8/3/2018: We have added a PyTorch implementation of an SRNN layer that is a drop in replacement for nn.LSTM. This layer has been minimally tested and is under development.

Matlab implementation for Skipping Recurrent Neural Networks from:

Learning Visual Storylines with Skipping Recurrent Neural Networks
Gunnar A. Sigurdsson, Xinlei Chen, Abhinav Gupta
http://arxiv.org/abs/1604.04279

The code is a MATLAB implementation of a Recurrent Neural Network, wrapped by the S-RNN architecture. This code was recently released, so please let me know if you encounter any strange behaviour.

The code is organized as follows:

• rnn_trainnet.m is a script used for training the S-RNN
• rnn_testnet.m is a script used for generating a summary of multiple albums, and selecting the best summary
• predictnext_short.m is a code to run the short-term prediction experiment from the paper
• predictnext_long.m is a code to run the long-term prediction experiment from the paper

The subset of Yahoo Flickr 100M used in the paper: https://www.dropbox.com/s/ghrieaikaobukz4/storylines_data.zip?dl=1 (9.4 GB)

This contains albums with image URLs and fc7 features in the format used by the code.

https://www.dropbox.com/s/66cal9wa0iumgoa/evaluation.zip?dl=1 (0.4 GB)

Using the below models should result in the following numbers corresponding to columns in Fig 6. right side in the paper.

• srnn: 0.310648
• rand: 0.214815
• nn: 0.2875

Pre-trained models: https://www.dropbox.com/s/zpuvm436ortrsw6/srnnmodels.zip?dl=1 (0.02 GB)

Please cite the following if it helps your research!

@article{sigurdsson2016learning,
  author = {Gunnar A. Sigurdsson and Xinlei Chen and Abhinav Gupta},
  title = {Learning Visual Storylines with Skipping Recurrent Neural Networks},
  journal = {European Conference on Computer Vision},
  year = {2016},
  url = {http://arxiv.org/abs/1604.04279},
  poster = {http://www.eccv2016.org/files/posters/P-3A-26.pdf},
  pdf = {http://arxiv.org/pdf/1604.04279.pdf}
}