Delete ToDo.md

ToDo.md

-- READ THIS ALL AND DELETE DUPLICATES & DEPRECATED(S) 
-
-- do we split the synthetic datasets?
-- keep it simple -> only go crazy-shiny for big conferences/journals!!!
-- reduce the num of imgs to 2-3k on each dataset and fully test a loop with the 4 datasets (from init to report)!!!
-- check kitti_to_yolo vs eval (validity, etc.),
-- auto highlight/textbf latex!!
-- check bdd100K txt vs png vs sorting,
-- deal with missing labels the same way on both datasets!!
-- yaml, etc.
-- run an experiment for each yalm in configurations | setup(s) -> one per dataset # decoupled solutions !!!
-- change the Bdd preprocessor so that you don't need the dummy gt fields!
-- figure out why some labels seem to be missing,
-- fix the (first) images with some obj but with empty labels!
-   - dropping them might be fine for our FP use case but NOK for PR guarantees!
-- enforce bounds & constraints satification,
-- max additive should be more valid?
-- replace the dummy detector by dummy datasets & a pass through/identity detector?
-- verify the number of images that are used for fine-tunning, it seems too small!!! -> the split params!!!
-- add a single seed everywhere untill it all becomes fully repeatable!!!
-- we use the accessible 80k not 100k
-- make it repeateable, fold-preemptable and fold-auditable,
-- compare this code base with your other/cleaner codebases and apply the discovered improvements (#FileByFile),
-- look at the data and chose the best semi-parametric copulas (ex: Gaussian for synth) using Kendall Tau, etc.
-- replace residues by residuals in your text,
-- try Zero hessians as that (de/reactivated) warning advices,
-- maybe, access the yaml with get & default values?
-- rename the detector for synthetic datasets or just add the dts name to the "Logging results" ... msg?
-- rename Split(s) as Fold(s) everywhere (#FileByFile)?
-- in demo, auto figure out if temp or not, which dataset, detectors, etc.
-- in future codebases 
-   -> foreach dts then foreach detector (instead of what done here!!!!)
-   -> add log files to your experiments -> better search txt than scroll cmd lines!!!
-- merge this to do list with the others!
-- new best practice = set gloabl seeds and local/dedicated rnd generator for each stocha op using the same utils.SEED 
-- stop manual testing!
-- figure out weather using jpgs instead of pngs could speed up or lighten things!
-- use multiple devices,
-- (re)implement reporting with a single axis and a single table per dataset!!!
-- finalize the README.md file!
-- figure out why DE hangs after convergence > 1.0
-- add the handling of folds with zero (or one) size, i.e., simple splits! 
-- remember to put back the right values in exp...yaml
-- add the pdf on github, 
-- check the formatting of your final version (editors, etc.),
-- make your code compatible with using multiple devices!
-- /!\ save the config in results.pkl!!!
-- fully run zhang et al on the gpu -> accelerate it!!
\ No newline at end of file