Darwin-Project

An advance algorithm for depression and anxiety detection from extracted facial and audio features. The implementation is based of several AVEC challenge participants with a number of personal improvements and additions.

Setting up the Environment

To create the environment run:

conda env create -f darwin_env.yaml

Then activate it with

conda activate darwin

If you install any new libraries, make sure you do it with conda unless there is no other choice. Then, update the environment with

conda env export --from-history > darwin_env.yaml

Running Experiments

To run you own experiment you have to first edit the config.ini file paying careful attention to naming conventions and ensuring you are not duplicating already existing functionality with a different name.

1. Set a name for experiment
2. For a single model:
   2. Use appropriate model name (Should be same as function name in model.py)
   3. Set model weights to 1
   4. Use appropriate name for features
   5. var_ratio determines how many components to keep after PCA
3. For multiple models:
   1. Separate the models and there attributes by a + sign
   2. Make sure model weights add to 1
4. If model has not been added yet, create a new function in model.py and add a call to it in the switcher statement. Make sure to use try except clause inside you function to avoid errors.
5. If features don't exists add them to relevant feature script and then load them in trough data.py functionality. Make sure the final data is formatted correctly.
6. Run your experiment and look up the results by running mlflow ui in same directory as mlruns folder

Optimizing your model

You can optimize your code by running pipelie.optimize and selecting parameter appropriately. Use a tuple to determine min and max for each of the desired parameters. For example (100, 1000) will search for the best parameter in range between 100 and 1000. The optimization uses a Bayesian Optimization technique and more information can be found here.