Parallel Stacking Classifier

Implementation of parallel computing stacking classifier using Message Passing Interface. Stacking classifier is based on 4 classifiers:

• RidgeClassifier
• RandomForestClassifier
• LinearDiscriminantAnalysis
• GaussianNB

Parallel computing workflow:

Master process sends the data to the slave processes, every process classifies using bounded for this process algorithm and then sends the classification results back to the master process.

Requirement

• mpi
• python 3
• virtual-env

How to run ?

1. Activate Virtualenv. Instruction

2. Install dependecies:

pip install -r requirements.txt

3. Run:

Note that you need to specify parameters like -t (type of run), -m (method), -d (dataset name)

Example invocation:

python -m parallel-stacking-classifier -t parallel -m test-train -d MNIST

Available types:

• sequence (runned by python)
• parallel (runned by mpiexec)

Available methods:

• CV
• test-train

Available datasets:

• MNIST
• CIFAR-10
• CIFAR-100
• letter-recognition

program will automatically create process using formula: 1 physical core = 1 process if you want manually specify number of processes add flag -n (--numberOfProcesses)

Example:

python -m parallel-stacking-classifier -t parallel -m test-train -d MNIST -n 2

Note that number of processes can be choosen only for parallel program invocation, the same invocation for sequencial flow will not work

program runned with module mpi4py to avoid deadlocks ( the finalizer hook of mpi4py module will call MPI_Abort() on the MPI_COMM_WORLD communicator, thus effectively aborting the MPI execution environment.)