Muffnn

Changed

- Increase minimum version of `tensorflow` to v1.15.4 to fix the security vulnerability reported in https://github.com/advisories/GHSA-4g9f-63rx-5cw4 (105)
- Set more specific `scikit-learn` version requirements to avoid incompatibilities and test failures (105)

Removed
- Remove Python 3.5 support (105)

Fixed
- Fixed an issue with builds failing due to numerical issues (103)

Changed
- Increase minimum version of `tensorflow` to v1.15.2 to fix the security vulnerability reported in https://github.com/tensorflow/tensorflow/security/advisories/GHSA-977j-xj7q-2jr9 (101).
- Dropped support for Python 2.7 and 3.4 (101).

Changed
- Allowed a recent version of `scikit-learn` (99).

Fixed
- Updated tests for changes in new versions of `scipy`, `scikit-learn`, and `flake8` (98).
- Increased required version of `tensorflow` due to published CVEs in older versions (98).

Added

- Added `prediction_gradient` method for understanding the impact of different
features in MLPs with dense inputs.
- Added support for SELU activations with alpha dropout.
- Added sample weights for the `FMClassifier`.
- Added `FMRegressor`.

Fixed

- Exposed `muffnn.__version__`.
- Fixed bug in `FMClassifier` where it failed for predicting one example.
- Fixed ValueError for type of target in `MLPClassifier` and `FMClassifier` (90).

Changed

- Updated requirements on numpy to 1.14 or higher.
- Updated requirements on scipy to 1.0 or higher.

Added

- Added support for the `sample_weight` keyword argument to the `fit`
method of MLPClassifier and MLPRegressor (75).

Changed

- Switched from requiring TensorFlow 1.x to 1.4.x because 1.5.0 was causing
Travis CI failures with Python 3.6 (78).

Added

- Added a `transform_layer_index` keyword and `transform` method to the
MLPClassifier and MLPRegressor to extract features from a hidden layer (62).

Changed

- Moved the MLPClassifier and MLPRegressor to using
[Xavier initialization](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/xavier_initializer) (68).