David Nicholson

In theory the config allows instantiating multiple models for training / prediction but this was never fully implemented and to make it really work is more trouble than it's worth...

ENH: Convert post-processing to Transform classes that can be specified in config

1

e.g. MajorityVote ```toml [vak.model.TweetyNet] [vak.model.TweetyNet.output-transform.MajorityVote] [vak.model.TweetyNet.output_transform] MinSegmentDur = {min_dur = 2} ``` :thinking: not sure if this is valid TOML, but ... something like this

e.g. https://github.com/keras-team/governance/blob/master/keras_api_design_guidelines.md#design-end-to-end-workflows-not-individual-functions-and-classes write the workflows first, then refactor around those what would a "through the Python API" workflow look like? what would "testing multiple models at once" look like?

Engine + core/cli re-design

4

edit: this issue was originally about considering backends but I am hijacking it to collect my thoughts on how to (re)design everything ---------------------------------------------------------------------------------------- it would really be nice to not...

allow data_dir option in PREP section to be a list of directories

2

instead of forcing user to e.g. write a bash script with a bunch of commands + separate .toml files

set up CI for GPU functionality with dvc-cml?

2

probably with an approach like this https://cwpearson.github.io/post/20190520-self-hosted-azp/ here's the associated repo: https://github.com/cwpearson/azure-pipelines-agent edit: this used to be about using Azure but thinking it would be better to run "on prem"...

automate updating vak-test-data repo on osf with datalad-osf

1

possibly with https://github.com/cognoma/figshare or https://github.com/rmcgibbo/figshare

adapted from https://github.com/yardencsGitHub/tweetynet/tree/master/article/src/scripts/demo-gif for example

ENH/CLN: replace sub-classing Engine/Model methods with callbacks / classes

1

we want to [avoid subclassing for code sharing](https://hynek.me/articles/python-subclassing-redux/) WRT this, `engine.Model` should still be used for training / predicting / etc., but - the pseudo-private `_train`, `_predict` etc should be...

DOC: add how to use built-in annotation formats

1

with demos that are tested, e.g. using TextGrid and Audacity datasets I get the feeling people are now assuming they *have* to convert to `'simple-seq'` Also we should have the...