globus-labs / cascade / 14757320726

95%

push

github

Multi-head replay finetuning (#67)

* Start interface for finetuning with replay

- Tool for pulling relevant frames out of initial dataset
- Start the datamodel for passing replay instructions

* Add routines for pulling replay data

* Bug fix: ensure entires are Python floats

* Introduce function for creating models with shared layers

* Introduce replay into training function

* Add cli args for freezing layers

* Add replay to the training CLI script

* Indentation issue: was adding another replay each epoch

* Use pytorch-ignite's logic for skipping epochs

* More identation issues.

What was I thinking!?

81 of 85 new or added lines in 4 files covered. (95.29%)

982 of 1030 relevant lines covered (95.34%)

0.95 hits per line