Skip to content

Latest commit

 

History

History
63 lines (55 loc) · 3.01 KB

changelog.md

File metadata and controls

63 lines (55 loc) · 3.01 KB
Raw

Changelog

Versions are tagged in git under v0.x

Version 0.5

  • able to compare not only with one specified, but also with others -> that should probably be a command line tool (after training)
  • distributions into the same plot to make them more comparable -> histograms: old vs new
  • add gradient clipping -> should be added
  • add categories in config
  • add linting
  • add bitbucket pipeline to run tests automatically

Version 0.4

  • plot creation for already trained weights -> write cli where weights are loaded
  • only train on the first reco track of a track
  • make an average loss graph from all experts -> how to communicate with the other experts? -> easy cause we use threads -> use class where all log to and if all have logged for an epoch we can log to tensorboard and create visualizations -> problems as they are somewhat in different processes and therefore cant really communicate, solve with shared memory maybe
  • baseline model v2 with BN and Relu
  • reweighting of trainings sample, by duplicating samples per bin or by reweighting them per bin -> or random sampling with same prob. per bin, idea: make classification problem
  • train with different batchsizes and learning rates per expert
  • write readme page
  • add unit tests
  • create pickle file with z, theta predictions after training for future comparision
  • dont use dataset predictions but optionally the ones from older trainings
  • file with single output number -> over all experts and per expert and maybe compare to previous
  • distribution sampling with in config: distribution should be configurable
  • organize main better and support cmd args
  • global experiment log
  • filter functions
  • native filter datasets
  • add presentation and finish readme
  • fix inhereting bug in config
  • add cli parameter which can overwrite config

Version 0.3

  • add statistical values such as mean and std to the plots (in form of legends)
  • implement rprop, generalize optimizers and put them into config
  • fix weight init
  • act function into the config
  • fix x axis of hist plot for gt data
  • add diff hist plot -> z(Reco-Neuro)
  • add std(z(Reco-Neuro)) to tensorboard plot metrics and relative old vs new
  • add std bins plot
  • pin pytorch lightning version
  • rescale z/theta outputs to represent real physical values
  • reimplement dataset caching
  • in the end of the training create weights, predication dataset, plots as pngs, and maybe evaluate test?
  • validate with best trained data
  • export (the best) weights in the end of the training

Version 0.2

  • description in experiment log
  • add extending of other config
  • add training for only z component
  • add easy dict
  • save git diff and commit id into log
  • add change log
  • per expert hparams -> which overwrite the default value e.g. different bach sizes
  • issue with singal shutdown in thread
  • models and critic function also in configuration using dicts