globus-labs / cascade / 9898805368

95%

push

github

Implement a serial version of learning on the fly (#22)

* Implement calculator method for ANI

* Add an example dynamics class

* Summarize how to use the class

* Change terminology

* Do not re-add data to the end of traj

* Started on implementing proxima

* Complete the unit test, fix bugs while at it

* Fix how we calculate the alpha parameter

The paper might have it backwards :|

* Make it so alpha will always be positive

Don't fit a intercept

* Add utilities for saving state of otf learner

* Add an option to reduce training frequency

* Add demonstration notebooks

* Test that state is serializable

* Handle edge case where len(data) % batch_size = 1

Ugh

* Threshold should be _lower_ than guess to start

* Ensure original object is not altered

* Ensure logger works

It is not engaged if we .step instead of .run

* Increase the weight for forces

1:10 is what I've seen in papers, and we've had it 1:0.9

* Ravel matrices before appending to list

* Turn example notebook into CLI script for better experiments

* Improve tests, add todos

125 of 128 new or added lines in 4 files covered. (97.66%)

482 of 513 relevant lines covered (93.96%)

0.94 hits per line