Training Classifiers using fastai and inference using TorchScript

This repository contain files for training classifiers using fastai and inference using TorchScript. Fastai is a great library, it provides a great APIs build on top of PyTorch like databunch and learner for creating dataloaders and training scheduler respectively. It uses lot's of state of the art methods for training and data creation like cyclic learning rates with One Cycle Policy, lot's of augmentations, etc. As we know, Python is slow as compared to other languages like C++, Java, etc. When we use our models in production, we need lesser inference time so, there using Python is not a very good idea, so we can use C++. Recently, PyTorch introduced TorchScript, an intermediate representation of a PyTorch model (subclass of nn.Module) that can then be run in a high-performance environment of C++. This repository contains on of the best methods to training a classifier and inference in TorchScript.

It contains several files and notebooks which can be used for specific purposes:-

1. notebooks/classifier.ipynb - This is a very well commented notebook for training a classifier in fastai.
2. notebooks/torchscript.ipynb - This is a notebook to convert PyTorch model to torchscript model(a serialized intermediate representation). In this notebook, I have shown how to convert a function to torchscript form which then can be used in C++.
3. classifier.cpp - This is a script written in C++ for inference of classifier. In this script, we are using model which we saved using notebooks. Here, we are using OpenCV for reading image and preprocessing.
4. CMakeLists.txt - cmake file for building and running the classifier written in C++.

Requirements

    Python>=3.6
    torch >= 1.4                                                                                                             
    OpenCV >= 4.1
    torchvision >= 0.4
    fastai >= 1.0

Intsructions for using the code

For using this project you need to follow some steps:

1. First clone this repository in your system and then navigate to repository folder

   git clone https://github.com/adityak2920/Traning-and-Inference.git
   cd Traning-and-Inference
   

2. Create a folder build

   mkdir build && cd build
   

3. Before procedding further we need to download latest distribution of libtorch to build our application which you can download from here. After downloading, unzip the folder and your libtorch directory should looks like this:

   libtorch/
           bin/
           include/
           lib/
           share/
           build-hash
           build-version
   

4. Now run these commands to build the application

   cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch ..
   cmake --build . --config Release
   

5. Now you can run your app with(you can specify name of app in cmake file)

     ./your_app
   

6. From next time you can recompile and run your app using

   make
   ./your_app
   

For further information regarding this you can learn from following sources:-
1. fastai course v-3
2. Intro to TorchScript
3. Loading a TorchScript Model in C++

Thanks to PyTorch Forums for resolving my doubts in implementing this project.