search

tticoin / LSTM-ER

Implementation of End-to-End Relation Extraction using LSTMs on Sequences and Tree Structures in ACL2016.
Apache License 2.0
225 stars 73 forks source link

Exit unexpectedly at training #4

Closed tijiang13 closed 7 years ago

tijiang13 commented 7 years ago

When running this code after compiling, the program gets killed at the first iteration of training.

build/relation/RelationExtraction --train -y yaml/parameter-semeval-2010.yaml

306560 words, 200 dimensions
start loading documents in data/semeval-2010/corpus/train/
start loading documents in data/semeval-2010/corpus/test/
==== iteration: 1 =====
train: 
killed

After adding some print, I find it breaks when conducting the function

# LSTM-ER/relation/RelLSTMModel 
# void RelLSTMModel::predict(vector<Table*> &tables, bool do_update, bool output, bool ent_only)
sgd_->update(params_.gradient_scale()/params_.minibatch()); # the 306 line

and the output of build/relation/RelationExtraction --train -y yaml/parameter-semeval-2010.yaml shows that this program treat every case as negative, which is obvious wrong.

Do you have any idea? Or have you tested your code in a brand new (virtual) machine with Ubuntu?

mmiwa commented 7 years ago

The model is not trained correctly, and maybe the error is due to the memory limit (process is killed by the OS).

As in README.md, we have developed the system with Fedora. We once succeeded to run it on Ubuntu and we have got several working reports, so this program should work on Ubuntu.

tijiang13 commented 7 years ago

Thanks, I originally run this code in a virtual machine, and it seems that the memory allocated by default is merely 4GB, which turns out to be far from enough.

tijiang13 commented 7 years ago

CLOSED.