SurpriseProbing:

A simple set of scripts to extract Wav2vec2 loss based on sequences generated by the ControlledStim dataset (https://github.com/PierreOrhan/ControlledStim). Make sure to have install the ControlledStim dataset:

    cd yourpath/ControlledStim
    pip install -e ./

This repository was developed purely for replication purposes and is not meant to be extended.

The ANN package subclass Huggingface implementation of Wav2vec2, correcting for a few issues. The probe package provides the loss.

System requirements: Same hardware and software requirements as the Transformers python package.

Software requirements: The software was tested in torch 2.0.1. Make sure to have torch < 2.0.1 as the loading of the transformers Wav2vec2 models might otherwise fail. After torch installation following the official repository instructions, install the transformers, matplotlib, tqdm amd zarr packages.

Installation:

    cd yourpath/StructureProbing
    pip install -e ./