Bpnet-lite

==============

Highlights
----------

- Fixed issue where non-linear operations in DeepLiftShap were not
registered correctly and hence causing minor divergences. Through the
use of an ugly wrapper object this has been fixed.
- Added in `print_convergence_deltas` and `warning_threshold` to the
`calculate_attributions` function and the `DeepLiftShap` object. The first
will print convergence deltas for every example that gets explained and the
second will raise a warning if the divergence is higher than it.

==============

Highlights
----------

- Extended support for the `chrombpnet` command-line tool
- Now has mirrored functionality of the `bpnet` command-line tool
- `chrombpnet pipeline` now mirrors `bpnet pipeline` except that it will
run each of the reports on each of the three models: the full ChromBPNet
model, the accessibility model, and the bias model. It will train a bias
model and an accessibility model if not provided.
- Changed the ChromBPNet object to be compatible with the `bpnet` command
options.
- Fixed issue with attributions where performance would degrade over time.

==============

Highlights
----------

- Extended support for the `bpnet` command-line tool
- Added in `marginalize` command-line option for generating those reports
- Added in `pipeline` command-line option for running a full pipeline from
model training to inference, attribution, tfmodisco, and marginalization

==============

Highlights
----------

- I forgot.

==============

Highlights
----------

- Addition of a `ChromBPNet` model
- Addition of an explicit, shared, `Logger` class
- "Peak" semantics have been switched to "locus" semantics


chrombpnet.py
-------------

- Newly added.
- This file contains the `ChromBPNet` class, which is a wrapper that
takes in two BPNet objects: a pre-trained bias model, and an untrained
accessibility model, and specifies the training procedure for training
the accessibility model.


io.py
-----

- The semantics of "peaks", e.g. `extract_peaks`, has been changed to loci,
e.g. `extract_loci`, and the associated keywords (now `loci` from `peaks`)
can take in a list or tuple of files to interleave them. This means you
can now train on peaks and background regions.


logging.py
----------

- Newly added.
- This file contains the `Logger` class which is a simple way to record
and display statistics during training.