neubig / modlm Goto Github PK
View Code? Open in Web Editor NEWmodlm: A toolkit for mixture of distributions language models
License: GNU Lesser General Public License v3.0
modlm: A toolkit for mixture of distributions language models
License: GNU Lesser General Public License v3.0
I try to run example but the execution of process.sh
results in "Segmentation fault".
The contents of example/result-ptb/mkn+w1x200-lstm200x1-sent-adam-mb0512-0.001_o-1-s100.log
are as follows.
/opt/modlm/example/../src/modlm/modlm-train --cnn_seed 100 --cnn_mem 0512 --wildcards "all aa ab ac ad ae af ag ah ai aj" --vocab_file model-ptb/ptb.train.vocab --train_file data-ptb/ptb.held.WILD.txt --valid_file data-ptb/ptb.valid.txt --test_file data-ptb/ptb.test.txt --dist_models "model-ptb/ptb.train.WILD.ngram_mkn_4_3_2_1 model-ptb/ptb.train.all.uniform model-ptb/ptb.train.all.unk" --learning_rate 0.001 --rate_decay 1.0 --model_dropout_prob 0.0 --model_dropout_decay 1.0 --dropout_models "" --trainer adam --training_type sent --clipping_enabled true --layers lstm:200:1 --node_dropout 0.0 --word_hist 1 --word_rep 200 --penalize_unk false --epochs 1000 --online_epochs -1 --max_minibatch 0512 --whiten "" --whiten_eps 0.001 --heuristic "" --evaluate_frequency 1 --model_out result-ptb/mkn+w1x200-lstm200x1-sent-adam-mb0512-0.001_o-1-s100.mod 2>&1 >> result-ptb/mkn+w1x200-lstm200x1-sent-adam-mb0512-0.001_o-1-s100.log &
[cnn] random seed: 100
[cnn] allocating memory: 0512MB
[cnn] memory allocation done.
Started training! (s=7.51e-07)
Reading vocabulary... (s=8.7075e-05)
Started reading model model-ptb/ptb.train.all.ngram_mkn_4_3_2_1 (s=0.00298226)
Started reading model model-ptb/ptb.train.all.uniform (s=5.57657)
Started reading model model-ptb/ptb.train.all.unk (s=5.57661)
Finished reading models (s=5.57662)
Creating model (s=5.57663)
Creating data for data-ptb/ptb.valid.txt (s=5.57856)
Creating data for data-ptb/ptb.test.txt (s=5.82293)
Started reading model model-ptb/ptb.train.aa.ngram_mkn_4_3_2_1 (s=6.09834)
Creating data for data-ptb/ptb.held.aa.txt (s=11.2268)
Started reading model model-ptb/ptb.train.ab.ngram_mkn_4_3_2_1 (s=11.8082)
Creating data for data-ptb/ptb.held.ab.txt (s=16.8796)
Started reading model model-ptb/ptb.train.ac.ngram_mkn_4_3_2_1 (s=17.4329)
Creating data for data-ptb/ptb.held.ac.txt (s=22.7245)
Started reading model model-ptb/ptb.train.ad.ngram_mkn_4_3_2_1 (s=23.0686)
Creating data for data-ptb/ptb.held.ad.txt (s=28.345)
Started reading model model-ptb/ptb.train.ae.ngram_mkn_4_3_2_1 (s=28.6992)
Creating data for data-ptb/ptb.held.ae.txt (s=33.9831)
Started reading model model-ptb/ptb.train.af.ngram_mkn_4_3_2_1 (s=34.308)
Creating data for data-ptb/ptb.held.af.txt (s=39.6078)
Started reading model model-ptb/ptb.train.ag.ngram_mkn_4_3_2_1 (s=39.9576)
Creating data for data-ptb/ptb.held.ag.txt (s=45.2559)
Started reading model model-ptb/ptb.train.ah.ngram_mkn_4_3_2_1 (s=45.5813)
Creating data for data-ptb/ptb.held.ah.txt (s=50.8563)
Started reading model model-ptb/ptb.train.ai.ngram_mkn_4_3_2_1 (s=51.1832)
Creating data for data-ptb/ptb.held.ai.txt (s=56.4738)
Started reading model model-ptb/ptb.train.aj.ngram_mkn_4_3_2_1 (s=56.8594)
Creating data for data-ptb/ptb.held.aj.txt (s=62.1225)
Done creating data. Whitening... (s=62.7943)
--- Starting epoch 1: online, lr=0.001 (s=63.2522)
./process.sh: line 47: 10561 Segmentation fault (core dumped) $MDIR/src/modlm/modlm-train --cnn_seed $SEED --cnn_mem $MEM --wildcards "all aa ab ac ad ae af ag ah ai aj" --vocab_file model-$SUF/$DATA.vocab --train_file data-$SUF/$HELDW.$LANG --valid_file data-$SUF/$VALID.$LANG --test_file data-$SUF/$TEST.$LANG --dist_models "$MODELS" --learning_rate $LR --rate_decay $RD --model_dropout_prob $DROPOUT_PROB --model_dropout_decay $DROPOUT_DECAY --dropout_models "$DROPOUT_MODELS" --training_type $TRAINING_TYPE --trainer $TRAINER --clipping_enabled $CLIP --layers "$RECURRENCE:$NODES:$LAYERS" --node_dropout $NODE_DROPOUT --word_hist $WHIST --word_rep $WREP --penalize_unk false --epochs $EPOCHS --online_epochs $OEPOCHS --max_minibatch $MINIBATCH --whiten "$WHITEN" --whiten_eps $WHITEN_EPS --heuristic "$HEURISTIC" --evaluate_frequency $EVAL_FREQ --model_out result-$SUF/$LOGID.mod
I'm using dependent libraries of the following revisions.
cnn revision: b39e16e2879df7f8043b192caaaf71b8d58cd78e
Eigen revision: 9546
Thanks!
The README says to use the "v2" branch of dynet. As far as I can tell, there is no branch of dynet called "v2". I can't get this repo to build using the master branch of dynet.
I'm trying to figure out how 's are handled during training and evaluation. There is a flag called "penalize_unk" that can be set to true or false. If it is true then I expect s to be treated the same as any word in the vocabulary. If it is false then I expect the 's to be ignored during training and testing.
The first issue is that there is no way to remove unk's from the n-gram models. This means that these will always assign probability mass to the unk token even when we are telling it that there are no unk's in the test data.
The second issue is that when the penalize_unk flag is false, the way the average log likelihood seems to be calculated is to not include in the numerator but to still include it in the denominator. If there are a lot of unks then the perplexity becomes really low. What I would expect is that the unk's should not be counted in both the numerator and the denominator. I'm pretty sure this is what SRILM does. This behavior is non-intuitive and makes me wonder why the penalize_unk flag is set to false in the example script.
I'm still working on understanding all the code. If I can figure it out myself I will try and submit a fix myself.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.