249
75%

Ran 13 Dec 2018 10:49AM UTC

Jobs 1

Files 77

Run time 5s

Badge

Embed ▾

pending completion

Build # 249

Build Type

push

travis-ci

Committed by

michelole

Commit Message

Properly truncate short narratives during prediction

When training, we learn `maxLength`, which is the largest number of sentences found on a single narrative.

When predicting, we initialize a feature vector with the same size as the document to be predicted *or* `maxLength`, whatever is *smaller*. This avoids filling the feature vector with invalid data in the case of a too short document, thus leading to a wrong prediction. This also avoids initializing a feature vector which is longer than the longest training sample.

This improves n2c2 accuracy as such:
Training: 66.45% -> 96.08%
Test: 68.34% -> 74.42%

Furthermore, training accuracy now correctly fits the stop criterion of 0.5, as evaluated by the dl4j library.

This was reported upstream on deeplearning4j/dl4j-examples#761.

Run Details

1344 of 3002 relevant lines covered (44.77%)

0.45 hits per line

Jobs

ID	Job ID	Ran	Files	Coverage
1	249.1	13 Dec 2018 10:49AM UTC	0	44.77	Travis Job 249.1

Source Files on build 249

Detailed source file information is not available for this build.

bst-mug / n2c2 / 249
75%

README BADGES
x

Markdown

Textile

RDoc

HTML

Rst

Jobs

Source Files on build 249

bst-mug / n2c2 / 249 75%

README BADGES x

Markdown

Textile

RDoc

HTML

Rst

Jobs

Source Files on build 249

bst-mug / n2c2 / 249
75%

README BADGES
x