bxshi / dynamic_sampled_softmax_loss

A new TensorFlow op that calculates sampled softmax loss with dynamic number of true classes per training instance.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

TensorFlow op for dynamic number of true classes per instance.

Usage:

Make sure you have TensorFlow installed, and if you are using an virtualenv, source your virtualenv first before compile the code.

cmake .
cmake --build .

You will get a .so or .dylib under the project root depending on your OS.

An example and test cases can be found in dynamic_candidate_sampling_example.py.

A modified version of this is used in ConMask. If you use this code, please consider cite Shi B., Weninger T., Open-World Knowledge Graph Completion, AAAI 2018. Thank you!

About

A new TensorFlow op that calculates sampled softmax loss with dynamic number of true classes per training instance.


Languages

Language:Python 62.1%Language:C++ 36.4%Language:CMake 1.4%