dmol-book
dmol-book copied to clipboard
New Chapter: hyperparameter selection
- [ ] Initial draft
- [ ] Applied example
https://docs.ray.io/en/latest/tune/index.html
a bit more far-fetched would be experiment tracking.
I'm thinking of stuff like weights and biases which 1) has tools for hyperparameter sweeps 2) tools to visualize some chemistry
as framework, I really enjoy using Optuna
@whitead Do you want to use this package (https://docs.ray.io/en/latest/tune/index.html) in this chapter?
@mehradans92 I think that comment was me sharing some existing methods used. It would be better to be as package agnostic as possible though.
@mehradans92 Read through it briefly. Looks great, a lot of work went into it! Also I can tell it will be very helpful. A few proposed changes:
- [ ] Try to look at the layers chapter once more, there is some overlapping material (e.g., dropout, regularization, hyperparameters).
- [ ] Cite some papers on learning rate schedulers and maybe add some information on momentum, since it's related. Also some have mentioned warm-start, which I'm not familiar with. Maybe mention it..
- [ ] Fig 8.2 - does it need to be a movie? Can be distracting while reading. I can see the benefit for 8.1 certainly
- [ ] Batch size - would love to get 1-2 citations here on batch size and its connection to randomness in estimating gradient
- [ ] Dropout - can you cite the paper and maybe add a bit more on where it should be added (all layers?), if it should be combined with other regularization, etc.
- [ ] It is really critical to use validation data for a hyperparameter search - otherwise you're implicitly fitting to testing data. See here. You need to strongly emphasize this point early and make sure code/examples uses the word validation, instead of test, for the search.
- [ ] On Keras, can you reduce the output level of the logging (verbose=0) so the text isn't rendered in the chapter.
- [ ] You've split up the code nicely, but it'd be great to have some discussion, maybe showing how snippets of how the methods work too, before going right into training.