Various places in the reservoir code assume the last dim of input arrays is the feature dimension, which is not the case when the input is a 2D field. The training code calls the assure_same_dims function which adds a z dim of size 1 if other 3D fields are present, but this doesn't handle the edge case where only 2D inputs are present (this would probably be the format of data for SST models). At inference, the adapter assumes the last dim of each variable is the feature dim. For both of these cases we should add some code early on in the array creation to reshape to add a feature dim of size 1 for 2D fields.
Various places in the reservoir code assume the last dim of input arrays is the feature dimension, which is not the case when the input is a 2D field. The training code calls the
assure_same_dimsfunction which adds a z dim of size 1 if other 3D fields are present, but this doesn't handle the edge case where only 2D inputs are present (this would probably be the format of data for SST models). At inference, the adapter assumes the last dim of each variable is the feature dim. For both of these cases we should add some code early on in the array creation to reshape to add a feature dim of size 1 for 2D fields.