DLEC - Deep Learning Emotion Classifier

Description

DLEC is a convolutional neural network that classifies images of human facial expressions in emotions.

It can be used as a framework for developing general purpose deep learning classifiers.

Motivation

This is a final year project to achieve the Bachelor degree in Computer Science Lutheran University of Brazil.

Objectives

Accomplish researches on deep learning and develop a model capable to classify human emotions in images.

How to use

Configuring the framework

train.py

• Set the train_dataset path
• Set the validation_dataset path (if empty, the train dataset will be slited into train and validation)
• Set image_width and image_height
• Set learning_rate, test_size (only used when validation_dataset is None), batch_size and epochs

cnn.py

• Define the architecture of the CNN in the function get_network_architecture
• Use the inline documentation as reference

predict.py

• Set the same image_width and image_height as in train.py

Training

• Make sure the images exist under the paths set in train_dataset and validation_dataset
• The images of each class should be in its own directory. See the configuration below:

     Datasets
     /      \
  Train   Validation
   /           \
Class 1,      Class 1,
Class 2,      Class 2,
  ...,          ...,
Class n,      Class n,

• Run the training with the command python train.py
• The training status will be outputed during training

Predicting

• Run python predict.py to get help on the available prediction methods

• Examples:
  • python predict.py 1 /Datasets/Test/ (evaluates the model)
  • python predict.py 6 (real time predictions using the embedded camera)
  • python predict.py <TASK> <PATH> (general case)

• Option 1 does not require the task specification. The command python predict.py <PATH> will run into task 1
• Option 6 does not need any path, since it will use the embedded camera as source

Technical notes

• The code is runnable both in Python 2 and Python 3
• The code also works on Windows (it was tested on Windows 10). Keep in mind adjustments might be required for your environment.

Author

Maikel Maciel Rönnau
Computer Scientist
maikel.ronnau@gmail.com
Linkedin - GitHub