search

Sujit-O / pykg2vec

Python library for knowledge graph embedding and representation learning.
MIT License
602 stars 109 forks source link

KG2E crashes after around 400 epochs #183

Open ArkDu opened 3 years ago

ArkDu commented 3 years ago

I run train.py with python train.py -mn kg2e -exp true -device cuda. It first seemed alright, but after around 400 epochs it's mini-test results became NaN or 0.000000.

Screen Shot 2020-09-13 at 8 46 50 PM

This is kg2e_Test_results.csv after training finished. It indicates that the model stopped working at around 400 epochs

baxtree commented 3 years ago

Hi, @ArkDu, I noticed the default learning rate for KG2E is 0.01, which may have been too high and resulted in a large gradient during the training. Would you like to try decreasing it and see if that helps?

ArkDu commented 3 years ago

Sure I'll try a smaller learning rate and see how it goes.