An extensible PyTorch framework to experiment with neural-networks-based deep learning algorithms on multiple data modalities for binary classification. The code in this repository was written by @matteo-rizzo for his Master's thesis.
- Dynamic multi-modal architectures. Off-the-shelf and easy-to-combine basic architectures (e.g. CNN, GRU, etc…) for multimodal approaches to binary classification tasks;
- Structured experiments. Run multiple reproducible iterations of cross-validation using different random seeds or metadata and generate detailed reports (including metrics, plots and predictions);
- Data management. Ready-to-use grouped data splits for K-fold, generated anew or based on metadata;
- Ease of extension. Clear extension points and easy customisation for different use cases.
Main high-level packages:
classifier. The core of the project, including the fundamental classes handling the dataset, the training and evaluation of the models as well as the neural networks architectures;preprocessing. A pipeline-structured utility that allows for the generation and preprocessing of the data. For instance, image representations of eye-tracking data can be produced from the corresponding sequences, while the latter need to be adjusted and polished before being fed to a model;utils. Various utility scripts for post-processing the results (e.g. aggregating the scores produced by multiple iterations of CV).