Repository files navigation

This repository is the out project about mood recognition using convolutional neural network for the course Seminar Neural Networks at TU Delft.

67% Accuracy

We use theFER-2013 Faces Database, a set of 28,709 pictures of people displaying 7 emotional expressions (angry, disgusted, fearful, happy, sad, surprised and neutral).The dataset quality and image diversity is not very good and you will probably get a model with bad accuracy in other applications!

You have to request for access to the dataset or you can get it onKaggle. Downloadfer2013.tar.gz and decompressfer2013.csv in the./data folder.

Install all the dependencies usingvirtualenv.

virtualenv -p python3 ./source ./bin/activatepip install -r requirements.txt

The data is in CSV and we need to transform it using the scriptcsv_to_numpy.py that generates the image and label data in thedata folder.

$ python3 csv_to_numpy.py

By default this is using AlexNet architectures, but in the paper we propose different ones.

# To train a model$ python3 emotion_recognition.py train# To use it live$ python3 emotion_recognition.py poc

Link