GC - Machine Learning

Table of Contents

1. Directory
2. Development Rules
3. Train Models
4. ML API
5. Helpful Flags For API
6. FAQ
7. Pull Requests

Directory

├── gamechangerml
│   ├── api
│   │   ├── README.md
│   │   ├── __init__.py
│   │   ├── docker-compose.override.yml
│   │   ├── docker-compose.yml
│   │   ├── fastapi
│   │   ├── getInitModels.py
│   │   ├── kube
│   │   ├── logs
│   │   ├── tests
│   │   └── utils
│   ├── configs
│   ├── corpus
│   ├── data
│   │   ├── features
│   │   │   ├── abbcounts.json
│   │   │   ├── abbreviations.csv
│   │   │   ├── abbreviations.json
│   │   │   ├── agencies.csv
│   │   │   ├── classifier_entities.csv
│   │   │   ├── combined_entities.csv
│   │   │   ├── corpus_doctypes.csv
│   │   │   ├── enwiki_vocab_min200.txt
│   │   │   ├── generated_files
│   │   │   │   ├── __init__.py
│   │   │   │   ├── common_orgs.csv
│   │   │   │   ├── corpus_meta.csv
│   │   │   │   └── prod_test_data.csv
│   │   │   ├── popular_documents.csv
│   │   │   ├── topics_wiki.csv
│   │   │   └── word-freq-corpus-20201101.txt
│   │   ├── ltr
│   │   ├── nltk_data
│   │   ├── test_data
│   │   ├── training
│   │   │   └── sent_transformer
│   │   ├── user_data
│   │   │   ├── gold_standard.csv
│   │   │   ├── matamo_feedback
│   │   │   │   ├── Feedback.csv
│   │   │   │   └── matamo_feedback.csv
│   │   │   └── search_history
│   │   │       └── SearchPdfMapping.csv
│   │   └── validation
│   │       ├── domain
│   │       │   ├── query_expansion
│   │       │   ├── question_answer
│   │       │   └── sent_transformer
│   │       └── original
│   │           ├── msmarco_1k
│   │           ├── multinli_1.0
│   │           └── squad2.0
│   ├── mlflow
│   ├── models
│   │   ├── ltr
│   │   ├── msmarco_index
│   │   ├── qexp_20211001
│   │   ├── sent_index_20211108
│   │   ├── topic_models
│   │   └── transformers
│   │       ├── bert-base-cased-squad2
│   │       ├── distilbart-mnli-12-3
│   │       ├── msmarco-distilbert-base-v2
│   ├── scripts
│   ├── src
│   │   ├── featurization
│   │   │   ├── abbreviation.py
│   │   │   ├── abbreviations_utils.py
│   │   │   ├── extract_improvement
│   │   │   ├── generated_fts.py
│   │   │   ├── keywords
│   │   │   ├── make_meta.py
│   │   │   ├── rank_features
│   │   │   ├── ref_list.py
│   │   │   ├── ref_utils.py
│   │   │   ├── responsibilities.py
│   │   │   ├── summary.py
│   │   │   ├── table.py
│   │   │   ├── term_extract
│   │   │   ├── test_hf_ner.py
│   │   │   ├── tests
│   │   │   ├── topic_modeling.py
│   │   │   └── word_sim.py
│   │   ├── model_testing
│   │   ├── search
│   │   │   ├── QA
│   │   │   ├── embed_reader
│   │   │   ├── query_expansion
│   │   │   ├── ranking
│   │   │   ├── semantic
│   │   │   └── sent_transformer
│   │   ├── text_classif
│   │   ├── text_handling
│   │   └── utilities
│   ├── stresstest
│   ├── train

Development Rules

• Everything in gamechangerml/src should be independent of things outside of that structure (should not need to import from dataPipeline, common, etc).

Configs

• Config files go in gamechangerml/configs. When you add a new class, import it in gamechangerml/configs/init.py.
• File paths in gamechangerml/configs/* should be relative to gamechangerml and only used for local testing purposes. Feel free to change on your local machine, but do not commit system specific paths to the repository.
• A config class (i.e., from gamechangerml/configs/*) should not be required as an input parameter to a function. However, a config class attribute can be used to provide parameters to a function (foo(path=Config.path), rather than foo(Config)).

What Can Be Stored On GitHub?

• Models and large files should NOT be stored on Github.
• Data should NOT be stored on Github, there is a script in the gamechangerml/scripts folder to download a corpus from s3.

Use Best Practices

• Code should be modular, broken down into smallest logical pieces, and placed in the most logical subfolder.
• All classes, functions, etc. should have clear, concise, and consistent docstrings.

  • Function docstrings should include:

    • A short description
    • Any important remarks
    • Parameter types, defaults, and descriptions
    • Return types and descriptions

    Example:

    def say(words, loud=False):
      """Make the animal say words.
    
      Args:
        words (str): Words for the animal to say.
        loud (bool): True to make the animal say the words loudly, False to 
          make the animal say the words in a normal tone. Default is False.
    
      Returns:
        None
      """

• Include a maximum of 1 class per file.
• Include README.md files that contain what, why, and how code is used.

Getting Started

To use gamechangerml as a python module

• pip install .
• you should now be able to import gamechangerml anywhere python is available.

Train Models

1. Setup your environment, and make any changes to configs:

• source ./gamechangerml/setup_env.sh DEV

2. Ensure your AWS enviroment is setup (you have a default profile)
3. Get dependencies

• source ./gamechangerml/scripts/download_dependencies.sh

4. For query expansion:

• python -m gamechangerml.train.scripts.run_train_models --flag {MODEL_NAME_SUFFIX} --saveremote {True or False} --model_dest {FILE_PATH_MODEL_OUTPUT} --corpus {CORPUS_DIR}

5. For sentence embeddings:

• python -m gamechangerml.train.scripts.create_embeddings -c {CORPUS LOCATION} --gpu True --em msmarco-distilbert-base-v2

ML API

1. Setup your environment, make any changes to configs:

• source ./gamechangerml/setup_env.sh DEV

2. Ensure your AWS enviroment is setup (you have a default profile)
3. Dependencies will be automatically downloaded and extracted.
4. cd gamechangerml/api
5. docker-compose build
6. docker-compose up
7. visit localhost:5000/docs

Helpful Flags For API

• export CONTAINER_RELOAD=True to reload the container on code changes for development
• export DOWNLOAD_DEP=True to get models and other deps from s3
• export MODEL_LOAD=False to not load models on API start (only for development needs)

FAQ

• I get an error with redis on API start
  • export ENV_TYPE=DEV
• Do I need to train models to use the API?
  • No, you can use the pretrained models within the dependencies.
• The API is crashing when trying to load the models.
  • Likely your machine does not have enough resources (RAM or CPU) to load all models. Try to exclude models from the model folder.
• Do I need a machine with a GPU?
  • No, but it will make training or inferring faster.
• What if I can't download the dependencies since I am external?
  • We are working on making models publically available. However you can use download pretrained transformers from HuggingFace to include in the models/transformers directory, which will enable you to use some functionality of the API. Without any models, there is still functionality available like text extraction avaiable.

Pull Requests

Please provide:

1. Description - what is the purpose, what are the different features added i.e. bugfix, added upload capability to model, model improving
2. Reviewer Test - how to test it manually and if it is on a dev/test server. (if applicable) i.e. hit post endpoint /search with payload {"query": "military"}
3. Unit/Integration tests - screenshot or copy output of unit tests from GC_ML_TESTS_119, any other tests or metrics applicable