About freeze

[lou-yimin]

Tranks fro your work.
Does the LIFT result in the paper come from a model where the backbone is frozen?

[MogicianXD]

When the backbone is small (e.g., DLinear), we jointly train LIFT and a randomly initialized backbone.

As for PatchTST, Crossformer, and MTGNN, most results are based on a frozen backbone. Empirically, there is little difference between the two training schemes. We recommend building LIFT on a pretrained and frozen backbone so as to save time costs.