feat: ship the per-track means and the target scaling, so predictions can be converted to signal units
#2
by Yanisadel - opened
Adds what a user needs to turn the predictions back into signal units. The model predicts in the scaled space it was trained on, and until now nothing in the repository said how that space relates to the signal.
New keys in track_metadata (both config.json and track_metadata.json, existing keys unchanged):
track_means_per_file_assembly: the mean of each track, per assembly, in the same order asbigwigs_per_file_assemblyrna_seq_per_file_assembly: whether each track gets the RNA-seq powertarget_scaling: the transform and its inverse
The model card gets a short paragraph with a copyable to_signal function.
These are the values the post-training used (get_scaling_params, mean_non_zero of the track statistics, RNA-seq when the assay contains "rna-seq"), identical at the four commits of the post-training chain. With them, the predictions over the figure 2C windows match the paper's to Pearson >= 0.999 (human, pig).
Yanisadel changed pull request status to merged