AlexNet Inference Using CUDA

This repository provides a CUDA-based implementation of AlexNet for performing inference on input images using the CUDNN library. It loads pre-trained weights from ONNX models (not included in this repository) and is based on NVIDIA’s CUDA sample code for LeNet, adapted for AlexNet architecture.

Overview

This project implements the AlexNet model for image classification, optimized for GPU-based inference using CUDA and cuDNN libraries. The implementation includes custom layers, fully connected and convolutional layers.

**Link to pre-trained model weights: **: ONNX AlexNet Model Weights

Requirements

To use this code, ensure the following dependencies are installed:

• CUDA Toolkit (Version 10.1 or newer)
• cuDNN Library (for accelerated GPU processing)
• OpenCV (for image processing)

This project has been tested with CUDA Compute Capability sm_75 but should work on any compatible device with minor adjustments.

Installation

1. Clone the repository:

   git clone https://github.com/Po-V/AlexnetCUDNN.git
   cd AlexnetCUDNN
   

2. Set up weights: Download the ONNX pre-trained weights for AlexNet in binary format amd store them in model_weights folder.

3. Compile and run the project:

   make
   ./alexnet
   

TODO

• Add support for fp16