tensorflow_speech_recognition_solution
tensorflow_speech_recognition_solution copied to clipboard
Fix : Semi-Supervised Learning fails
Problem
Current implementation fails to reproduce semi-supervised
method you mention in solution.pdf
.
This is due to data.py
#L103~104 where all the test data is assigned a label of len(self.label_words_dict)
, which is unknown
.
There is no execution error, but kaggle submission results in ~0.15.
Solution
- return both list of data and list of label in
get_sub_list()
so that correct labels are fed for semi-supervised setting - this fix results in public/private leaderboard score ~0.85 which is decent, but does not improve overall ensemble score as mentioned in
solution.pdf