An exact asynchronous algorithm for distributed optimization over digraphs based on this paper.
The codes are tested on Linux (Ubuntu 16.04) with OpenMPI and Python 3.6.
- To run a demo on a least-square problem, execute the following command
mpiexec -np $(num_nodes) -bind-to core -allow-run-as-root python ./asyspa/asy_gradient_push.pyThe -bind-to core is optional, which tells the MPI to bind each process to a core, and hence the num_nodes should not be larger than the number of cores in your machine.
The command above will distributedly solve the following problem using the AsySPA
where the data is randomly generated.
- To distributedly train a multi-class logistic regression classifier on the Covertype dataset as in the paper, run the following command first
python ./asyspa/preprocess_dataset_covtype.py --n $(num_nodes) --file ./dataset_covtype/covtype.csvThis will preprocess the data and divide it into num_nodes parts. Then, run the following command to start distributed training
mpiexec -np $(num_nodes) -bind-to core -allow-run-as-root python ./asyspa/distributed_asy_logistic_regression.py --data_dir ./dataset_covtype/data_partition_$(num_nodes) --save_dir ./result/core_$(num_nodes)This will train the classifier for 300 seconds in default, and you can change the time by passing some arguments. See the file for details.