@tacazares The train.md documentation looks great! Minor suggestions
random ratio default should be 0 (as used in the final models). Please reference peak-centric training, so people can figure out how the parameter relates to what's discussed in the manuscript.
what are other options for output activation? (I see there's a traget_scale_factor, so I'm assuming l2-loss might be an option?)
I wasn't sure I understood what --shuffle_cell_type flag was, is this pan-cell training option? If so, would be helpful to mention so. Also, for best performance, default should be true.
Might be worth including a figure legend for the beautiful figure. You could mention words like peak centric, pan-cell, random ratio, so people connect the dots.
@tacazares The train.md documentation looks great! Minor suggestions