Switch branches/tags
Nothing to show
Clone or download
Latest commit 41ae790 May 13, 2018

Emotion recognition with CNN

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

Angry Test

67% Accuracy

Angry Test


We use the FER-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 on Kaggle. Download fer2013.tar.gz and decompress fer2013.csv in the ./data folder.

Install all the dependencies using virtualenv.

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

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

$ python3

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


# To train a model
$ python3 train
# To use it live
$ python3 poc