lrjconan / rbp

Recurrent Back Propagation, Back Propagation Through Optimization, ICML 2018

RBP

This is the PyTorch implementation of Recurrent Back Propagation as described in the following ICML 2018 paper:

@article{liao2018reviving,
  title={Reviving and Improving Recurrent Back-Propagation},
  author={Liao, Renjie and Xiong, Yuwen and Fetaya, Ethan and Zhang, Lisa and Yoon, KiJung and Pitkow, Xaq and Urtasun, Raquel and Zemel, Richard},
  journal={arXiv preprint arXiv:1803.06396},
  year={2018}
}

Setup

To set up experiments, we need to build our customized operators by running the following scripts:

./setup.sh

Dependencies

Python 3, PyTorch(0.4.0)

Run Demos

• To run experiments X where X is one of {hopfield, cora, pubmed, hypergrad}:

  python run_exp.py -c config/X.yaml

Notes:

• Most hyperparameters in the configuration yaml file are self-explanatory.
• To switch between BPTT, TBPTT and RBP variants, you need to specify grad_method in the config file.
• Conjugate gradient based RBP requires support of forward mode auto-differentiation which we only provided for the experiments of Hopfield networks and graph neural networks (GNNs). You can check the comments in model/rbp.py for more details.

Cite

Please cite our paper if you use this code in your research work.

Questions/Bugs

Please submit a Github issue or contact [email protected] if you have any questions or find any bugs.

rbp's Issues

Run with updated packages

pyenv install 3.6.0
pyenv shell 3.6.0
python3.6 -m venv ~/.virtualenvs/rbp
source ~/.virtualenvs/rbp/bin/activate
wget https://download.pytorch.org/whl/torch-0.4.0-cp36-cp36m-linux_x86_64.whl
# Select matching torchvision from https://pypi.org/project/torchvision/
wget https://download.pytorch.org/whl/torchvision-0.4.0+cpu-cp36-cp36m-manylinux1_x86_64.whl
pip install torch-0.4.0-cp36-cp36m-linux_x86_64.whl
pip install torchvision-0.4.0+cpu-cp36-cp36m-manylinux1_x86_64.whl
pip install numpy matplotlib cffi sklearn pyaml easydict
bash setup.sh
python run_exp.py -c config/hopfield.yaml

Gives

Traceback (most recent call last):
  File "run_exp.py", line 47, in <module>
    main()
  File "run_exp.py", line 15, in main
    config = get_config(args.config_file)
  File "/home/tsbertalan/Dropbox/Projects/RBP/utils/arg_helper.py", line 38, in get_config
    config = edict(yaml.load(open(config_file, "r")))
TypeError: load() missing 1 required positional argument: 'Loader'

Installing instead the oldest pyaml I can with my pip in this env (pyaml==12.12.5) gives the same.

(rbp) tsbertalan@pico:~/Dropbox/Projects/RBP$ pip freeze
cffi==1.15.0
cycler==0.11.0
easydict==1.9
joblib==1.1.0
kiwisolver==1.3.1
matplotlib==3.3.4
numpy==1.19.5
Pillow==8.4.0
pyaml==12.12.5
pycparser==2.21
pyparsing==3.0.7
python-dateutil==2.8.2
PyYAML==6.0
scikit-learn==0.24.2
scipy==1.5.4
six==1.16.0
sklearn==0.0
threadpoolctl==3.1.0
torch==0.4.0
torchvision==0.2.2
tqdm==4.19.9
You are using pip version 9.0.1, however version 22.1.2 is available.
You should consider upgrading via the 'pip install --upgrade pip' command.

I know this code is somewhat old and probably a better approach is just to use it as a guide for a newer implementation, but would it be possible to at least provide a requirements.txt so I wouldn't waste time searching for and downloading incorrect old wheels? If not, I can update the pyaml usages later.

t-SNE visualization of hidden state trajectories

Hello @lrjconan! I just read your remarkable paper and loved the hidden state trajectory plot you included in the publication. Do you have any code available on how to make such a plot to allow to compare any other learning method to it? Or is this straight-forward to do as in t-SNE-ing all trajectories together, then plotting them as colored lines?