PASTIS/main_method/code/original_pretraining_source/README.md

Recovered original pretraining source

These are the supplied project's recovered pretraining/config/dataset/atrous-SRDA sources. The temporal encoder replaced the previous DaViT/Swin-style encoder; this is not a standard Swin backbone.

Run pretrain.py in this directory with PASTIS_ROOT set. Set EXP_ROOT to a new output directory so new pretraining does not overwrite existing weights. The notebook does this automatically.

Original pretraining:496fold-5patches,4patches/batch,sixteen32×32tiles per patch,1,984updates/epoch,100epochs,NDVI/frame-gated masked reconstruction. Source preserves historical comments; the current benchmark qualification supersedes old full-comparability wording.

Full notebook includes these files inline, checkpoint attachment, downstream adaptation and the complete final ensemble. The source is architecture-compatible with the supplied weights; re-running it is not a guarantee of reproducing every historical random or library detail.