adityasomak / pslqa

An implementation of Probabilistic Soft Logic Engine using Python/Gurobi

PSL Engine

This engine is created by building up on the theory presented in the Hinge-Loss Markov Random Fields and Probabilistic Soft Logic. Even though the authors have already published a open-sourced PSL engine, our version offers several advantages:

• The engine is purely pythonic which makes it easier to interface with many other NLP/Vision libraries.
• This engine is coded to directly integrate knowledge from ConceptNet and Word2vec. This is the engine that produced the required results in Explicit Reasoning over End-to-End Neural Architectures for Visual Question Answering which was presented in AAAI 2018.
• We use an off-the-shelf optimization library such as Gurobi engine to run inference and in our experience it is quiet fast.
• This engine especially built for Q&A. But can be extended to support generic rule-base written in PSL.

Citation

If you use or modify this engine for your project, please do not forget to cite:

@inproceedings{DBLP:conf/aaai/AdityaYB18,
   author    = {Somak Aditya and Yezhou Yang and Chitta Baral},
   title     = {Explicit Reasoning over End-to-End Neural Architectures for Visual Question Answering},
   booktitle = {Proceedings of the Thirty-Second {AAAI} Conference on Artificial Intelligence,
                New Orleans, Louisiana, USA, February 2-7, 2018},
   year      = {2018},
   crossref  = {DBLP:conf/aaai/2018},
   url       = {https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/16446},
   timestamp = {Thu, 03 May 2018 17:03:19 +0200},
   biburl    = {https://dblp.org/rec/bib/conf/aaai/AdityaYB18},
   bibsource = {dblp computer science bibliography, https://dblp.org}
}

Pre-requisites:

• Gurobi 6.5.0 or Gurobi 6.5.2 (please visit Gurobi Website for license and download information). Please note that academic licenses (multiple) are free.
• install gensim using pip install gensim to load word2vec models.
• install nltk using pip install nltk.
• Other packages to install: numpy, enum, fuzzywuzzy, sqlite3

For ConceptNet and Word2vec, download conceptnet-numberbatch-201609_en_word.txt and GoogleNews-vectors-negative300.bin and change the paths in W2VPredicateSimilarity.py.

To make changes and run from command Line:

Run the following commands:

• sudo python setup.py sdist
• sudo pip install --upgrade dist/PSLplus-0.1.tar.gz

---

To Run VQA-model Inference from command-line:

Use:

• python main.py vqa -pslcode <rules-file> -datadir <psl_test_data_dir> -parentDir <pslDataRootDir> -qaData <qaData> -option infer
• If you want to run the demo on vqa, use the test_pslqa.py under the vqa_demo/expt2_aaai18 directory:
  • python2.7 test_pslqa <qatestdir> <pslDataRootDir> <answerFile> -stage 2/3 -split test/dev <startFrom>
  • Arguments:
    • qatestdir: Expects generated Json files from DenseCaptioning software for each image under the subdirectory "densecap/"
    • pslDataRootDir: Stage 1 output will be stored here. Stage 2 output will be stored under "psl" subdirectory.
    • answerFile: List of possible top 1000 answers - one in each line.
  • Example: For running inference on dev split python2.7 test_pslqa vqa_demo/expt2_aaai18/densecap vqa_demo/expt2_aaai18/ <answerFile> -stage 3 -split dev 0
  • If you are able to run it successfully and want to produce desired results, play with the summation constraints (S_ANS and S_CAND) in core/pslground.py. For example, use 10.0 for both for the data in expt2_aaai18.

---

To Run generic-models Inference from command-line:

Use:

• python main.py generic -pslcode <rules-file> -datadir <psl_test_data_dir> -mention <argument_to_predict> -option infer
  • The mention argument indicates the engine what we want to append a summation constraint for this predicate. For example for the vqa the mention argument would be the ans_candidate.
• Note option learn is not implemented for generic pipeline.

To Run weight-learning from command-line:

• Coming Soon

Author Affiliation:

The code is developed by Somak Aditya, when he was working as a Graduate Research Assistant in Prof. Chitta Baral's lab at CIDSE, Arizona State Univerisity.

Disclaimer

• If you want to augment or fix an issue, please raise an issue in the Issue tab of the project. However, as I am the sole contributor, please note the answers will be infrequent. Please feel free to clone and modify based on your needs though. Cheers!
• If you are looking for all the functionalities of PSL 2.0 in one place, please visit the original groovy-based Gihub Repo.