Solves multicut (correlation clustering) problems orders of magnitude faster than CPU based solvers without compromising solution quality on NVIDIA GPU. It also gives lower bound guarantees.

We use CUDA 11.2 and GCC 10. Other combinations might also work but not tested. CMake is required for compilation.

git clone git@github.com:pawelswoboda/RAMA.git
cd RAMA
mkdir build
cd build
cmake ..
make -j 4

We also provide python bindings using pybind. Simply run the following command:

python -m pip install git+https://github.com/pawelswoboda/RAMA.git

We require multicut instance stored in a (.txt) file in the following format:

MULTICUT
i_1 j_1 cost_1
i_2 j_2 cost_2
...
i_n j_n cost_n

which corresponds to a graph with N edges. Where i and j should be vertex indices and cost is a floating point number. Positive costs implies that the nodes are similar and thus would prefer to be in same component and viceversa. Afterwards run:

./rama_text_input -f <PATH_TO_MULTICUT_INSTANCE>

An example to compute multicut on a triangle graph:

import rama_py
opts = rama_py.multicut_solver_options("PD")
rama_py.rama_cuda([0, 1, 2], [1, 2, 0], [1.1, -2, 3], opts) 

The solver supports different modes which can be chosen by initializing multicut_solver_options by following:

• "P": For running purely primal algorithm (best runtime).
• "PD": This one offers good runtime vs quality tradeoff and is the default solver.
• "PD+": For better quality primal algorithm (worse runtime).
• "D": For only computing the lower bound.

RAMA normally expects that:

1. Node indices always start from 0 and there are no missing node indices. For example on a graph with 1000 nodes, the node indices should be in [0, 999].
2. There are no duplicate edges.

If this is not the case with the input graph then use the option --sanitize_graph (for C++ CLI) and opts.sanitize_graph = True for python interface. Note that the output node labels would be computed for all nodes in the input graph (where nodes without edges will have label -1). Duplicate edges will be removed in a non-deterministic order. Also see this issue for more discussion.

The default set of parameters are defined here which correspond to algorithm PD from the paper. This algorithm offers best compute time versus solution quality trade-off. Parameters for other variants are:

• Fast purely primal algorithm (P): This algorithm can be slightly worse than sequential CPU heuristics but is 30 to 50 times faster.

  ./rama_text_input -f <PATH_TO_MULTICUT_INSTANCE> 0 0 0 0

• Better quality primal algorithm (PD+) : This algorithm can even be better than CPU solvers in terms of solution quality as it uses dual information. Still, it is 5 to 10 faster than best CPU solver.

   ./rama_text_input -f <PATH_TO_MULTICUT_INSTANCE> 5 10 5 10

• Dual algorithm (D): Use this algorithm for only computing the lower bound. Our lower bounds are slightly better than ICP and are computed up to 100 times faster.

   ./rama_text_input -f <PATH_TO_MULTICUT_INSTANCE> 5 10 0 0 5 --only_lb

Run ./rama_text_input --help for details about the parameters.

The above-mentioned Python solver takes input in CPU memory and then copies to GPU memory. In cases where this takes too much time we offer additional (optional) functionality in Python bindings which allow to directly use the GPU tensors and return the result in GPU memory. For this there are two options:

• Binding via pointers to GPU memory: Does not require compiling RAMA with PyTorch support (as done below). This option passes the GPU memory pointers to RAMA (the data is not modified). See test\test_pytorch_pointers.py for usage.

• Direct binding of Torch Tensors: To use this functionality ensure that PyTorch is built with the same CUDA version as the one used in this code and the ABI's match (see https://discuss.pytorch.org/t/undefined-symbol-when-import-lltm-cpp-extension/32627/7 for more info). Support for PyTorch can be enabled by:

WITH_TORCH=ON pip install setup.py

After this you should be able to run test/test_pytorch.py without any errors. To suppress solver command line output set opts.verbose=False.

If you use this work please cite as

@inproceedings{abbas2022rama,
  title={RAMA: A Rapid Multicut Algorithm on GPU},
  author={Abbas, Ahmed and Swoboda, Paul},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8193--8202},
  year={2022}
}