FEAT automatic transfer learning in consensus pipeline

[GeorgWa]

Allows automatic transfer learning of RT, charge, MS2 as part of the regular search workflow. Introduces a new config group to the .yaml.

transfer_learning:

  # if true, a custom peptdeep model will be created using the transfer learned library
  enabled: False

  # number of precursors per batch
  batch_size: 2000

  # maximum learning rate per batch.
  # The maximum learning rate will be reached after a warmup phase and decreased using a plateau scheduler
  max_lr: 0.0001

  # TODO remove and replaced by fixed 70:20:10 split
  train_ratio: 0.8

  # test every n intervals
  test_interval: 1

  # learning rate patience after which the lr will be halved
  lr_patience: 3

  # minimum precursor number to perform transfer learning
  minimum_psms: 10000

  # maximum number of epochs
  epochs: 51

  # number of warmup epochs during which the lr is ramped up
  warmup_epochs: 5

  # normalised collision energy encoded during training
  nce: 25

  # instrument type encoded during training
  instrument: 'Lumos'