tzw101 / cuda_engine_image_classifier_ Goto Github PK
View Code? Open in Web Editor NEWThis project forked from prathamesh-pawar/cuda_engine_image_classifier_
This project is aimed to convert a Tensorflow model built in Python into a Cuda Engine based on C++. This is done to reduce the latency of the code. Latency is a critical bottleneck for a lot of Deep Learning Projects. Majority of the tensorflow projects could not be materialized into real world solutions because the latency was impractical. However converting the trained model into a CUDA engine based in C++ could help us reduce the latency significantly. In this project we use fashion_MNIST dataset as an example. We train the dataset using a normal Tensorflow model and then save this model into an .uff file which is universally readable. We then use these weights to create a API based model in C++ and then feed this model the training images to calculate accuracy and most importantly latency.