Cranial Modeling

Generic tools

• Specialized iter-tools where each iterator can be used more than once, but does not load all data into memory.

Machine Learning Tools

• Standard model interface

  ...where "model" means any stateless or stateful transformation of data. Because of the standard interface each "model" can be used by itself or as a step in another "model". This is similar to scikit-learn pipelines, but at a lower level and with a use of specialized iter-tools instead of in-memory data.

• Standard scripts for deploying Models as microservices

  Including training, and using models for batch or online inference.

Models

Contains subclasses of ModelBase. These may have dependencies that must explicitly be included during pip installation.

Installation

The basic installation will include any dependencies needed if you are just writing your own models.

If you want to use soem of the pre-packaged models in cranial.models, you may need extra packages installed. Available extras are:

bpe-nlp Installs numpy in support of bpe and nlp. gensim Installs the gensim package to support gensim_models. p2p_utils Installs bs4 in support of p2p_utils. spacy Installs spacy in support or the spacy_tokenizer.

Example of installing spacy and bpe-nlp support

pip install "cranial[spacy, bpe-nlp]"

After spacy installation, the additional following step is required:

python -m spacy.en.download parser

Usage

(Work in progress. Tutorials forthcoming?)

1. Create a directory and change to it.
2. Create a file called model.py with a class called Model that inherits from cranial.model_base.ModelBase or cranial.model_base.StatefulModel.
3. Create a config.json file with at least {"model_name": "some_unique_name"}.
4. Run python3 -m cranial.service_scripts.serving --no-load.

About Cranial

Cranial is a Framework and Toolkit for building distributed applications and microservices in Python, with a particular focus on services delivering predictions from online learning models.

The machine learning components do not provide algorithms or models like SciKitLearn or Tensorflow or Spark or H2O, but instead provide wrappers so that models and pipelines created by these tools can be deployed and combined in standardized ways.