exalearn / ExaMol / 6935243578

98%

Build Type

push

github

Committed by web-flow

Commit Message

Support multi-fidelity active learning (#122)

* Start a base class for thinkers that use ML

* Move training logic into base class

* Start specification for multi-fidelity

* Move startup to the base class

* Make a "get or make" DB operation

* Include the inchi key method

* Add the task selection logic for multifi

* We are printing the InChI key now

* Keep track of which molecules are in the database

* Make PipelineThinker a subclass of Single

Was about to refactor in a way that the Single objective
thinker did not have any methods

* Fix the call to the super __init__

* Combined submit and store for inference

Will make it easier to subclass

* Refactor inference submission, implement multifi

* Fix model evicting logic

* Pass recipe along with task information

* Find levels given solution

* Filter out scores from search space molecules in database

Also start on multi-fidelity training

* Refactor to use new multifi scorer interface

E.g., pass lower fidelities as a kwarg

* Extend test for better coverage

* Made code more robust

1. Race condition of completed molecule being put back in queue
2. Molecules failing to parse during loading
3. One of two recipes not having desired fidelity

* Updated tests with new logging messages

* Only compute features once

* Document the new thinker

Run Details

284 of 293 new or added lines in 8 files covered. (96.93%)

2 existing lines in 1 file now uncovered.

2154 of 2197 relevant lines covered (98.04%)

0.98 hits per line