Current feature extraction strategy in AbstractTrainer is to extract feature for a story then downsample the feature according to user specified feature. However, user may want to use different features that are sampled at different rates, and downsample them with different downsampling strategies. For example, using static embedding features downsampled with a uniform filter and word rate downsampled with a lanczos filter (or not downsampled depending on implementation).
A solution is to move the downsampler into the feature extractor. This also has the benefit of ensuring the returned features match the number of TRs.
Current feature extraction strategy in
AbstractTraineris to extract feature for a story then downsample the feature according to user specified feature. However, user may want to use different features that are sampled at different rates, and downsample them with different downsampling strategies. For example, using static embedding features downsampled with a uniform filter and word rate downsampled with a lanczos filter (or not downsampled depending on implementation).A solution is to move the
downsamplerinto the feature extractor. This also has the benefit of ensuring the returned features match the number of TRs.