mlpack
blog
|
Augmented Recurrent Neural Networks - Week 5 + 6
Augmented Recurrent Neural Networks - Week 5 + 6
This two weeks featured a lot of weird bugs. The most memorable, in my opinion, are:
- The missing closing curly brace drove g++ crazy, which resulted in error message of length 27k lines (sic!). The link to the
make
log: gist - The copy-paste bug in
mlpack/models
repo. Because of that bug the LSTM model in all tasks was iterating the learning 20 times - even ifepochs
parameter was set to another value!
However, there was also several nice code refactorings:
- Automating the parameter check in
mlpack/models
. Unlike previous versions, now we use automated check for parameter validity and error message generation. - Templating the
RunTask
code inmlpack/models
. Now the task type is a template, unlike previous versions, where a separate function was crafted for every task.
Also, we keep on working on HAM implementation - for example, there is an open pull request on adding HAMUnit
, which already features tested implementation of tree data structure.
Finally, we have more or less stable experiment results for the three tasks:
- CopyTask: output log link
- AddTask: TBA
- SortTask: TBA
Generated by 1.8.13