blei-lab / edward / 4813

90%

push

travis-ci

Multinomial classification metrics (#743)

* evaluate kl-divergence for multinomial outputs (#664)

* add kl-divergence test (#664)

* if `total_count` > 1, generate single multinomial prediction before
evaluating metrics (#664)

* WIP: implement `compute_multinomial_mode`, add test

* WIP2: implement `compute_multinomial_mode`, add test

* WIP3: implement `compute_multinomial_mode`, add test

* WIP4: implement `compute_multinomial_mode`, add test

* if `output_key` is a {binary, categorical}_discrete, convert `probs` to tf.Tensor at end of if-block

* pep8 compliance

* add `multinomial_accuracy` metric, test

* ensure that `size` argument in np.random.choice is an int

* handle case when `total_count == 1`

* add scikit-learn-style binary classification aggregation metrics

* fix indents

* remove dummy file

* fix line-spacing

* expand dimensions of `y_true` and `y_pred` `if len(y_true.shape) == 0`

* correct 4-space indent to 2 spaces

* correct formula for kl-divergence

* implement API for specifying binary-averaging method; add tests

* add random seed to tests

* employ `assertAllEqual` when evaluating equality of numpy arrays

* fix pep8

* add docstrings

* cosmetic changes

* move softmax function into `compute_multinomial_mode`

* add docstring

* replace the red text

* ensure docstring lines < 80 characters

106 of 106 new or added lines in 4 files covered. (100.0%)

2979 of 3343 relevant lines covered (89.11%)

1.78 hits per line