Hyperbolic RNN in Pytorch

Prerequisites

You need Python3.6 to run the code. The list of the dependencies are in requrements.txt. Run:

python -m pip install -r requirements.txt

Training

To reproduce the results from the Table 1 on the final report, run the following command:

cd ./experiment_run
python run.py --data_dir=./data --num_epochs=30 --log_dir=./logs --batch_size=1024 --num_layers=2 --cell_type=hyp_gru

where you can change the argument cell_type=hyp_gru to cell_type=eucl_gru if you want to run Euclidean version of GRU.

Note that training can take up to 12-15 hours, having the batch size of 1024. You can increase the batch size to get the results more quicker, but wait for some slight accuracy drop.

Results

Model	Value
Fully Euclidean GRU / B=64	93.25
Fully Hyperbolic GRU / B=1024	96.8

TODO

• Abstract interface for Riemannian manifolds, embed- ded in ambient real coordinate space.
• Compatible generic RSGD.
• Compatible generic RAdam.
• Compatible implementation of Poincare ball and Mo- bius arithmetics.
• Test coverage for optimization routines.
• GRU based on Mobius arithmetics, API-compatible with torch.nn.GRU.
• Layers parameterized by pivots of Log and Exp, as opposed to fixed pivot of 0 in Mobius arithmetics-based layers.
• Test coverage for “Mobius” layers and RNN loops.
• Numerical stability with float64.
• Numerical stability with float32.
• Investigation of possibility of using cudnn loop.
• C++ implementation of core operations and loops.