This is a repository for Leo Kristopher Piel's Master's thesis. It contains code for the built models and conducted expoeriments.

NB! The code cannot be run, because the training, validation and test data are not included in this repository.

• baseline: contains files related to baseline system predictions analysis.
• dnn: contains all the saved feedforward dnn models. Also includes dnn.py predictions in file dnn_predicted_labels.npy that were used as labels for some of the trained RNN models.
• history: contains files with information about the training process of all the saved models.
• models: contains the saved state of the models. Models were mostly saved during the highest point of validation accuracy in training process. Two of the missing states can be found here:
  • gender_6.hdf5: https://drive.google.com/open?id=1qTjkdjSpG8GSOwSuHSb1wRH3p9twQcdK
  • model_22.hdf5: https://drive.google.com/open?id=1Gc_glXGtS-cmRvlO5e2sNtJaWggR0QE1
• predictions: contains models' predictions on validation and test data. Predictions ending with _males.py or _females.py are the ones of the models that were trained on gender specific data. Predictions ending with _males_original.py or _females_original.py are the ones of the models trained on the whole dataset.
• rnn: contains all the built RNN models
• survey: contains survey results in survey.csv and data analysis.

• History.ipynb: Jupyter Notebook of the analysis of the training process of the models. Creation of confusion matrices.
• Results.xlsx:
  • Sheet 1: accuracies of different models on validation and test data.
  • Sheet 2: Analysis of data division into age groups.
  • Sheet 3: survey results. Same as survey.csv in survey folder.
• combined_predictions_1.py: ensemble model, where best models for males' or females' age group identifciation was used after predicting gender.
• combined_predictions_2.py: ensemble model, where combination of the best models for males' or females' age group identification was used after predicting gender.
• merge_predictions.py: brings out all the results of different models and includes the ensemble models created for boosting the accuracies.