vocalpy/vak

A neural network framework for researchers studying acoustic communication

vak's Issues

change 'train_data_dict' and 'train_data_path' etc to just 'train_data'? in config / code

Less verbose, no confusion from slightly different names

apply `os.expanduser` to all paths in config

change config to attrs, write docstring for sub-config classes, use those in docs

add 'copy_training_data' argument for learncurve

add 'save_data' parameter to vak.cli.summary; add option to config

cleanly separate general `dataset.split` from more specific `train_test_dur_split`

see comments in #57

config parsing code throws bug when 'train_data_dict' not set even when running 'prep'

as discovered by @yardencsGitHub

Tensorflow 2.0 compatability

learncurve: grab all replicates and durations and only then run the training

to avoid crash halfway through training, as per @yardencsGitHub

add 'spectrogram_files' option to [DATA]

add contrib info + covenant to docs

adding detail here from @Luke-Poeppel
https://github.com/NickleDave/vak/issues/365#issue-906829778

It would be great to have a CONTRIBUTING.md file in the repo––or notes in the README––for new contributors to the package. This could contain info on:

1. Desired commit message formatting
2. Style guide
3. Testing

See here for a nice example.

Fix "experiments.md" in README.md

add ability for paths in VocalizationDataset to be relative

so that a vds generated in one place doesn't crash when used in another place.

Maybe add root or location attribute and have all others paths be relative to that?

change 'data' to 'prep'

Improve train/test(/val) split functions?

Making this into a general issue / discussion of how to improve these functions.
Will collect thoughts + ideas from other issues here (to close those and consolidate).

add separate config section for learncurve

add 'save_transformed_data' argument in docs

learncurve, train should use size of X_train, not what's in config

add DeepLabCut-style API

use crowsetta 2.0

add DeepChirp models as built-ins

https://github.com/kylerbrown/deepchirp

use tf.data, get rid of 'reshape_data_for_batching'

change 'n_syllables' to 'n_classes'

make Librosa a dependency, deprecate current spectrogram functions

remove `input_vec_size` arg from vak.utils.data.reshape_data_for_batching

and just use "height" of data

fix default None for [DATA] output_dir that gives confusing error

fix __repr__ for VocalizationDataset

currently dumps all attributes for every Vocalization in voc_list

Quick fix would be to override __repr__
Might be worth figuring out why attrs doesn't come up with a good __repr__ though

add 'annotation_format' and 'audio_format' options to [DATA]

change absolute imports to relative where possible

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

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

make it so invalid sections/options in config.ini raise human-interpretable errors

i.e. validate those sections and options

remove 'log_device_placement=True' from learncurve and summary

Not helping anymore + the major reason CI logs are too verbose

rename `get_inds_for_dur` to `subset_inds` and then write wrapper `subset`

that accepts X_train (stack of spects) and returns X_train_subset, using subset_inds

make 'make_spects_from_dir' func to wrap repeated logic from make_data/train/predict

one arg should be "purpose" or something like that with valid values {'learncurve','train','predict'} that then makes logic w/in function more explicit
i.e. if purpose=='predict' better than
Boolean flag that makes you invert logic mentally if is_not_predict==True

add `clip_size` option to data prep

throw better error if output section already has results_dir_made_by_main when running learncurve

should be some guide for contributing nets to library

something like sklearn
https://scikit-learn.org/stable/faq.html#new-algorithms-inclusion-criteria
to avoid a Pelican-like situation

add tqdm

document how `Makefile` works in dev install instructions

ENH: add ability to finetune / continue training

ENH: Add timing of training

I think the simple way to do this for now is just to log tic and toc before and after we call trainer.fit

unit tests that actually run and test the cli (not the functions that it calls)

fix how vak.main deals with --glob and --txt

can use combinatoric itertools to more efficiently draw subsets of training data?

i.e. find every combination of set of spectrograms with all classes present, then pick x of those

single source version

https://packaging.python.org/guides/single-sourcing-package-version/#single-sourcing-the-version

add 'all' command to cli, separate 'make_data' out of 'learncurve'

add spect, spect_ID_vector convenience methods to VocalizationDataset

fix predict

fix train so it works with validation set

accept generic Keras + Torch models instead of requiring user to subclass AbstractVakNetwork

less intimidating + confusing