Introduction

Hello everyone. At first, I trained ResNext50_32x4d to solve this problem, but I decided to train RegNet800. The motivation was to compare these two architectures and try the power of the network design space in practice . And as a result, RegNet shows itself better, so I'll leave the weights of this model as an estimate.

Setting up the environment

You can download all the necessary libraries, frameworks using requirements.txt

pip install -r requirements.txt

Train

Before launching train.py , run data.py . Since in this task I am doing data augmentation (because there are few of them).

data.py -> train.py

Eval

If you want to use the Simpsons classification model, then run the eval script.py and specify the path to your picture. And then the script will give you who is depicted in your picture.

launch eval.py -> enter path image

Docker

Instead of following all the steps described above, you can install Docker and pull my project from there. After installing Docker, run this command:

docker pull stolzor/classification_simpsons

After pulling my project, you need to run the following command:

docker run -d -p 5000:5000 stolzor/classification_simpsons

Complete! Now you only need to select the file to classify!