Parallel CNN

This project implements a Convolutional Neural Network (CNN) in C++ using multiple parallelization strategies:

• Sequential
• OpenMP
• MPI
• Hybrid (MPI + OpenMP)

The network is trained and evaluated on the MNIST dataset provided in CSV format.

---

Project Structure

parallel-CNN/
│
├── dataset/           # Contains mnist.zip which includes mnist_train.csv and mnist_test.csv
├── Sequential/        # Sequential C++ implementation
├── OpenMP/            # OpenMP parallelized CNN
├── MPI/               # MPI parallelized CNN
├── Hybrid/            # MPI + OpenMP hybrid CNN
└── README.md

---

Prerequisites

• C++ compiler:
  • g++ with OpenMP support
  • mpic++ (for MPI and Hybrid)
• Dataset:
  • Located in the dataset/ directory as a ZIP file

---

Dataset Setup

The dataset is provided as a ZIP archive: dataset/mnist.zip.

Before running any version of the code, unzip the dataset as follows:

cd dataset
unzip mnist.zip
cd ..

This will extract the following required files:

• dataset/mnist_train.csv
• dataset/mnist_test.csv

---

Compilation and Execution

1. Sequential Version

cd Sequential
g++ cnn.cpp -o cnn_seq
./cnn_seq

---

2. OpenMP Version

cd OpenMP
g++ -fopenmp cnn.cpp -o cnn_omp
./cnn_omp

To control the number of threads:

export OMP_NUM_THREADS=4

---

3. MPI Version

cd MPI
mpic++ cnn.cpp -o cnn_mpi
mpirun -np 4 ./cnn_mpi

Replace 4 with the desired number of MPI processes.

---

4. Hybrid Version (MPI + OpenMP)

cd Hybrid
mpic++ -fopenmp cnn.cpp -o cnn_hybrid
mpirun -np 2 ./cnn_hybrid

To control OpenMP threads per MPI process:

export OMP_NUM_THREADS=4
mpirun -np 2 ./cnn_hybrid

---

Output

All versions of the program will print:

• Training time
• Testing time
• Accuracy
• Confusion matrix
• RMSE