Skip to content

seraphlabs-ca/SentenceMIM-demo

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

57 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PWC

SentenceMIM-demo

This repo contains code to reproduce some of the results presented in the paper "SentenceMIM: A Latent Variable Language Model".

Code is based on: https://github.com/timbmg/Sentence-VAE. We included source code from npeet package for convenience (https://github.com/gregversteeg/NPEET).

Installation

git clone https://github.com/seraphlabs-ca/SentenceMIM-demo.git

Data is NOT included in repo (see below for downloading instructions).

You can install the required packages with the command below.

# requirements.txt
nltk==3.3
matplotlib==2.2.2
numpy==1.16.4
tqdm==4.28.1
torch==1.4.0
scipy==1.3.2
sklearn==0.0

NOTE: Virtual env is highly recommended to prevent corruption of your personal pip environment.

# Use at your own risk - virtual env is highly recommended
pip install requirements.txt 

TODO

  • Extend code to train on arbitrary text files
  • Add pre-trained models

Experiments

Below is summary of the results, and the corresponding commands to train and test MIM, VAE, and AE (with similar architecture) over Penn Treebank (PTB) dataset.

Dataset

Sentences Tokens Words per Sentence
Train Valid. Test Vocab. #w (max)
PTB 3370 3761 42068 10000 21 +/- 10 (82)

Penn Treebank (PTB) Dataset: Please download word-level data from https://catalog.ldc.upenn.edu/LDC99T42. Extract the files to data/datasets/ptb/. Expected files:

README
ptb.test.txt
ptb.train.txt
ptb.valid.txt

Results Summary

In what follows we summarize the results this code can produce. For qualitative results we show best performing MIM model (i.e., sMIM (512) ), and compare it to best performing VAE (i.e., sVAE (16) ), and to VAE with 512 dimensionality, sVAE (512), which suffers from collapsed posterior.

Perplexity

NLL/PPL results are an upper bound, computed with MELBO for MIM, and ELBO for VAE.

Model PPL NLL BLEU Parameters
AE (16) 0.2916 11M
AE (128) 0.5321 11M
AE (512) 0.5745 12M
sVAE (16) 113.14 107.38 0.0882 11M
sVAE (128) 117.2 108.19 0.0803 11M
sVAE (512) 121.7 109.04 0.0809 12M
sMIM (16) 75.29 98.14 0.2975 11M
sMIM (128) 28.58 76.14 0.5542 11M
sMIM (512) 19.02 66.9 0.6226 12M

AE is an auto-encoder with similar architecture (i.e., latent code z is taken to be mean of posterior q(z|x)). We present BLEU results only as it is not a PDF estimator.

Reconstruction

Below are reconstruction results for best performing models.

  • DATA - observation (i.e., input sentence)
  • MEAN RECON - reconstruction given the mean of the posterior q(z|x)
  • Z RECON - reconstruction given sample z from the posterior z ~ q(z|x)
  • Z PERT - reconstruction given sample z from a posterior with x10 larger variance

sMIM (512)

DATA: <sos> the system is the problem not an individual member
MEAN RECON: the system is not the problem an individual member <eos>
Z RECON: the system is the problem not an individual member <eos>
Z PERT: the system is not the problem an individual member <eos>


DATA: <sos> sony itself declines to comment
MEAN RECON: sony itself declines to comment <eos>
Z RECON: sony industry declines to comment <eos>
Z PERT: sony itself sony itself to n <eos>

sVAE (16)

DATA: <sos> the system is the problem not an individual member
MEAN RECON: they worry about their or ceiling at their own <unk> wellcome sears ' s not forget down <eos>
Z RECON: often con artists can ignore their credit judgments just how we had with them who now increase after <unk> losing <unk> <unk> and now discarded our country <eos>
Z PERT: stock-index arbitrage volatility volatility volatility volatility volatility portfolio portfolio portfolio portfolio portfolio volatility volatility portfolio portfolio portfolio portfolio portfolio portfolio portfolio portfolio portfolio portfolio portfolio portfolio portfolio ...

DATA: <sos> sony itself declines to comment
MEAN RECON: hyundai and other cities have been asked to exclude <eos>
Z RECON: brian <eos>
Z PERT: mcdonnell douglas turner turner turner pictures mother singer a genetic <unk> gene ogilvy former sanford <unk> turner pictures woman writer turner broadcasting university gene turner turner pictures <unk> <unk> university in genetic damage but her ted district was buddy and jazz <unk> in cells never complex kate judge o'kicki and certified mother <unk> turner pictures <unk> <unk> st . louis park <eos>

sVAE (512)

DATA: <sos> the system is the problem not an individual member
MEAN RECON: federal bank authority authority home loans to wedge about n n states at least n n increases <eos>
Z RECON: she ' ll remember her chaos he sees by a season that means <unk> and her sister in looking to dress video procedures to lift more confidence in the same time <eos>
Z PERT: initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated initiated ...

DATA: <sos> sony itself declines to comment
MEAN RECON: miniscribe said there are that there have been trimming <eos>
Z RECON: last week the 13th of israel gives the complaints for $ n million or n cents a share immediately <eos>
Z PERT: arabia arabia parity layer adjustable adjustable repeated repeated adjustable devise forum adjustable devise forum devise forum traffickers donuts forum devise screens repeated computerized repeated repeated repeated repeated screens repeated repeated screens screens screens repeated screens screens screens quina teachers screens screens repeated screens screens ...

Interpolation

Below are interpolation results for best performing models. Sentences in [ brackets ] are the source, and destination sentences.

sMIM (512)

[ <sos> the system is the problem not an individual member ]

the problem is not changing the an insider problem <eos>

the system is not the problem an individual problem <eos>

the problem is not the problem on an individual merits <eos>

the system system is not an individual member <eos>

the system itself is not an immediate <eos>

sony itself the problem is n't difficult to work <eos>

columbia corp . never put the voice to <eos>

sony corp . almost to comment the problem <eos>

sony itself declines almost to comment <eos>

sony itself declines to comment <eos>

[ <sos> sony itself declines to comment ]

sVAE (16)

[ <sos> the system is the problem not an individual member ]

all the programs and the people party held me how are set up leads and by boosting their earthquake he says <eos>

apparently she says the government ' s perfectly is born nervous that other <unk> chinese <unk> and lewis <unk> <unk> <unk> and psychiatric publishers such as brooks brothers brothers into leather ' s landmark the affiliates are an behavior for them <eos>

the combination they are bearing a new <unk> approach company can find anything that customers are n't registered of them by using the so-called global office to understand a balance of <unk> out of the information ' s health-care system <eos>

as much as n n it does n't think they should ask now from facing paying their work <eos>

one participant probably change in congressional policy was then at the same time the <unk> yesterday was sending by the <unk> ' s cat of management and his own colleagues who joined the work force of <unk> <eos>

<unk> is perfect in the state <eos>

this is as part of heart on the acquisition of the no . n marketplace hoped that everyone does n't believe that the british attorney reported a ruling at its forecast mr . <unk> has earlier a meeting with a company something that money he said in october <eos>

there is included that its number of bonds measures activity are still much less priced to yield to treasurys <eos>

tucson creating alcohol <eos>

the most form spacecraft reading in advertising of prime minister mr . watson <unk> the most dramatic attempt into a vast region <eos>

[ <sos> sony itself declines to comment ]

sVAE (512)

[ <sos> the system is the problem not an individual member ]

the five cardiovascular clearly sure her new line was unanimously as a signal that left money out after section n to join morrison <unk> rebel out at a hearing committee and chicago utility <eos>

at the same time the merksamer office ' s <unk> crisis <eos>

in the week the humans gives corporate cultural personnel even vicious and even acquisitions <eos>

<unk> economists like a source of his goodman and company showing about more than the <unk> <unk> <eos>

some atmospheric companies are urging products to meet with the <unk> image at the <unk> or inflated <eos>

as of the <unk> to become the <unk> mengistu mr . honecker said yesterday should be <unk> about the central bank <eos>

to tackle for mr . kasparov ' s patents tied to meet in sci tv said it would make his offer to pursue major business information is <unk> by a <unk> of cash <eos>

in the year-ago period goodyear added n n to n n <eos>

elizabeth r . <unk> and house <eos>

herbert hunt who follows the stock price <eos>

[ <sos> sony itself declines to comment ]

Sampling

Below are sampling results for best performing models.

sMIM (512)

SAMP: he expects meaningful retail margins for <unk> in terms of the previous obstacle about n to $ n million or customer world said reebok as <unk> corp . cleveland <eos>
                                                                                                                          
SAMP: <unk> old i ' ve made the computer maker to <unk> a higher complex settlement of arbitragers that were one problem <eos>

SAMP: our companies however cray computer financial institutions would n't discuss that the bank ' s proposed spending paid but when a character to one company with her <unk> inc . said
 <eos>                      

sVAE (16)

SAMP: at the strip and explain expectations will be discussed as n hours of production in the next day <eos>

SAMP: most of us agreed to recognize the techniques for the american market <eos>

SAMP: transportation serves and other minority interests have been eliminated in the absence fashion indeed the wage for greater doors <eos>

sVAE (512)

SAMP: most of recent problems still kept that most important contribution will follow unscrupulous when the sale of the premium will be because they have lately their money <eos>

SAMP: but they ca n't save the value of their drugs <eos>

SAMP: ford jaguar ' s net income fell n n to $ n million $ n million or $ n a share from $ n million <eos>

Training Commands

hidden_size=512
embedding_size=300
dataset=ptb
for latent_size in 16 128 512; do
    # MIM
    ./train.py  \
        --seed 1 \
        --test \
        --dataset ${dataset}  \
        --embedding_dropout 0.5 \
        --epochs 200 \
        --max_sequence_length 100 \
        --batch_size 20 \
        --embedding_size ${embedding_size} \
        --hidden_size ${hidden_size} \
        --latent_size ${latent_size} \
        --optim adam \
        -lr 0.001 \
        -x0 0 \
        --mim
    
    # VAE
    for x0 in 0 10000; do
        ./train.py  \
            --seed 1 \
            --test \
            --dataset ${dataset}  \
            --embedding_dropout 0.5 \
            --epochs 200 \
            --max_sequence_length 100 \
            --batch_size 20 \
            --embedding_size ${embedding_size} \
            --hidden_size ${hidden_size} \
            --latent_size ${latent_size} \
            --optim adam \
            -lr 0.001 \
            -x0 ${x0}
    done

    # AE
    ./train.py  \
        --seed 1 \
        --test \
        --dataset ${dataset}  \
        --embedding_dropout 0.5 \
        --epochs 200 \
        --max_sequence_length 100 \
        --batch_size 20 \
        --embedding_size ${embedding_size} \
        --hidden_size ${hidden_size} \
        --latent_size ${latent_size} \
        --optim adam \
        -lr 0.001 \
        --marginal \
        -x0 0
done

Testing Commands

NLL is upper bounded using MELBO for MIM, and ELBO for VAE.

./test.py \
    --max_sequence_length 100 \
    --test_epochs 1 \
    --seed 1 \
    --max_sample_length 20 \
    --batch_size 20 \
    --temperature 0.2 \
    --test \
    --split test \
    --test_bleu \
    --test_sample \
    --test_interp \
    data/torch-generated/exp/*

All results will be stored in the same path as the model, inside a text file best-test-1-marginal0-mcmc0.txt (name corresponds the command line arguments used with test.py script).

Citation

Please cite using the following bibtex entry

@ARTICLE{2020arXiv200302645L,
       author = {{Livne}, Micha and {Swersky}, Kevin and {Fleet}, David J.},
        title = "{SentenceMIM: A Latent Variable Language Model}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Computation and Language, Computer Science - Machine Learning, Statistics - Machine Learning, 68T50, I.2.7},
         year = 2020,
        month = feb,
          eid = {arXiv:2003.02645},
archivePrefix = {arXiv},
       eprint = {2003.02645},
 primaryClass = {cs.CL},
}

About

This repo contains code to reproduce some of the results presented in the paper "SentenceMIM: A Latent Variable Language Model"

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages