-
In order to manage weakly/unsegmented sequence learning tasks, Connectionist Temporal Classification (CTC) is an output Layer for compatible machine learning models, commonly recurrent neural networks (RNNs), such as LSTM.
-
It is introduced by Alex Graves, Santiago Fernández, Faustino Gomez and Jürgen Schmidhuber, and published in ICML 2006 with the tittle of Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks
-
Unfortunately, the original project was used to be working under Linux OS with the LSTM network and needs several dependencies.
-
This code is a refactored version of some part of the original code of the RNNLIB project by Alex Graves.
-
In this repository, file ctc.hpp is a modified version of the “TranscriptionLayer.hpp” (CTC Layer) file in the original project which does not require any internal or external (ex. Boost C++ Libraries) dependencies, so you can easily include and use it with any standard C++ compiler under any project on Windows or Linux.
- This library allows you to add CTC layer into your project without getting involved into handling inconsistency between dependency of your project. as an example combining CTC layer with the graphical models such as Conditional Random Fields (CRFs) family for utlizing them on Temporal/Unsegmented/Weakly Sequence Classification/Lablelling applications. For instance refer to my paper Training LDCRF model on unsegmented sequences using Connectionist Temporal Classification. This repository could easily be integrated into the legendary but orphan HCRF library project. 😁
- Boost data structure replace with standard C++ STL library.
- There is an option that allows you to easily modify the code to enable/disable the magical blank state!! for more information about this feature read an interesting article by Théodore Bluche that referenced in the Acknowledgments.
- Standard C++ 11+
- OS: Windows/Linux/Mac
Inspiration, references, etc.