Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Feature Request: default/auto parameters for training #33

Closed
chreul opened this issue Oct 24, 2018 · 1 comment
Closed

Feature Request: default/auto parameters for training #33

chreul opened this issue Oct 24, 2018 · 1 comment
Assignees
@ChWick

Comments

@chreul
Copy link
Member

chreul commented Oct 24, 2018

when training a model ensemble it would be nice if there was an "auto" option to automatically choose somewhat sane parameters, e.g. regarding early stopping. these parameters should depend on the number of GT lines available for training.
based on our experiments i propose the following auto defaults:

  • early_stopping_frequency: about half the number of available GT lines, maybe rounded up to the next hundred.
  • early_stopping_nbest: 5, i think 10 is too high. this is a general "issue" and not specific to this auto functionality.
  • max_iters: maybe 10 epochs? not sure about this one.
  • checkpoint_frequency = early_stopping_frequency.
@chreul chreul mentioned this issue Oct 24, 2018
Closed
ChWick added a commit that referenced this issue Oct 25, 2018
@ChWick
Switch default behavior of early stopping #33
af92971 
By default the checkpoint frequency is equal to the early stopping
frequency. If either one is set to -1 the other one will be used as
value.
ChWick added a commit that referenced this issue Oct 25, 2018
@ChWick
Enabling frequencies relative to epochs: #33
bb79ecf 
By default using an early stopping frequency of 0.5.
Checkpoint frequency uses by default early stopping frequency.
Display supports relative to epoch amount (0 < display <= 1).
ChWick added a commit that referenced this issue Oct 25, 2018
@ChWick
Automatical update of checkpoints
3af8e25 
Each checkpoints has a version. When accessing, old models with
automatically be upgraded to the current version.
Closes #27
Part of #33
@ChWick
Copy link
Member

ChWick commented Oct 29, 2018

  • As of bb79ecf: early stopping frequencies may be relative to epochs (if <= 1), default = 0.5.
  • early stopping number of best models set to 5
  • max iterations is unchanged, usually this parameter should never be required, because training with a validation set is recommended. Alternatively, the training can always be stopped manually.
  • By default the checkpoint frequency uses the early stopping frequency if not explicitly stated.

@ChWick ChWick closed this as completed Oct 29, 2018
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@ChWick ChWick

Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
2 participants
@ChWick @chreul