Skip to content

Latest commit

 

History

History
36 lines (27 loc) · 1.93 KB

README.md

File metadata and controls

36 lines (27 loc) · 1.93 KB
Raw

Neural Mixed Effect Models

This repository contains the evaluation code for "Neural Mixed Effects for Nonlinear Personalized Predictions" which is accepted at ICMI'23.

Please see here, if you are looking for only the Neural Mixed Effects (NME) code.

Setup

git clone [email protected]:twoertwein/NeuralMixedEffects.git
cd NeuralMixedEffects
poetry update  # installs dependencies

Usage

# NN-Generic
poetry run python ml/train.py --LINK <link> --LABEL <label>
# NN-Specific
poetry run python ml/train.py --LINK <link> --LABEL <label> --PER_PERSON --IND
# MLP-LME
poetry run python ml/train.py --LINK <link> --LABEL <label> --LME
# unregularized NME
poetry run python ml/train.py --LINK <link> --LABEL <label> --PER_PERSON
# NME
poetry run python ml/train.py --LINK <link> --LABEL <label> --NME

Where <label> can be: imdb, news, spotify, iemocapa (arousal on IEMOCAP), iemocapv (valence on IEMOCAP), valence (daily mood on MAPS), and constructs (the four affective states on TPOT).

<link> determines the model type where person-specific parameters are. It can be linear (last layer), all (all layers), first (first layer), crf (transition matrix T of CRF), linear+crf (last MLP layer and T), first+crf (first MLP layer and T), and all+crf (all MLP layers and T).

To reduce the training data for each person, please use --SMALLER 20 to train 20% less data per person.

Data

The features for Imdb, News, Spotify, MAPS, and TPOT are available here. If you want the features for IEMOCAP, please send us proof that you completed the data-sharing agreement required by IEMOCAP.