ResNet based Image classification using Pytorch C++

The project provides the model which is trained on ImageNet and can be used to classify the images. The project uses multi-thread approach to calculate the model predictions.

Run Using Docker (tested with Docker 20.10.7)

• Clone this repo.
• Pull the docker image environment using docker pull karthi0804/pytorch-resnet-cpp:deploy - it takes around 10 mins!
• Run the docker contianer using docker run --rm -it -v /absolute/path/to/repo:/classifier karthi0804/pytorch-resnet-cpp:deploy
• If the docker container is up successfully, you fill find the git code under the directory classifier inside the container. Please check by using root@XXXXXXXXXXXXXX:/# ls
• Go to project root using root@XXXXXXXXXXXXXX:/# cd /classifier inside the docker container.
• Make a build directory in the top level directory: mkdir build && cd build inside the docker container.
• Compile: cmake -DCMAKE_PREFIX_PATH=/libtorch .. && make inside the docker container.
• Run it: ./Pytorch-CNN-classifier inside the docker container.
• Modify the pic/ folder to add custom images.

Model Output

Input the num of workers: 2
Spawning workers...
Collecting results...
from worker : 140511188563712 : Top-1 Prediction with prob  71.2% of    ../pic/turtle.jpg: Label: box turtle, box tortoise
from worker : 140511196956416 : Top-1 Prediction with prob  97.1% of       ../pic/dog.jpg: Label: beagle
from worker : 140511188563712 : Top-1 Prediction with prob  23.9% of     ../pic/dog-1.jpg: Label: Cardigan, Cardigan Welsh corgi
from worker : 140511196956416 : Top-1 Prediction with prob  67.7% of     ../pic/shark.jpg: Label: tiger shark, Galeocerdo cuvieri
Inference took 2214 milliseconds

Code Structure

• main.cpp : contains the main code to create and call the class Inference.
• Inference : This class encapsulates the Torch Script module of ResNet along with other necessary fucntions like predict and display in inference.cpp
  • predict : splits the dataset and spawns multiple threads with each batch.
  • display : collects the model output from the threads and prints the Top-K predictions along with their probability.
• model.py : to generate torch script file.