Machine Learning Lifecycle Framework
Ebonite is a machine learning lifecycle framework. It allows you to persist your models and reproduce them (as services or in general).
pip install ebonite
Before you start with Ebonite you need to have your model. This could be a model from your favorite library (list of supported libraries is below) or just a custom Python function working with typical machine learning data.
import numpy as np def clf(data): return (np.sum(a, axis=-1) > 1).astype(np.int32)
Moreover, your custom function can wrap a model from some library. This gives you flexibility to use not only pure ML models but rule-based ones (e.g., as a service stub at project start) and hybrid (ML with pre/postprocessing) ones which are often applied to solve real world problems.
When a model is prepared you should create an Ebonite client.
from ebonite import Ebonite ebnt = Ebonite.local()
Then create a task and push your model object with some sample data. Sample data is required for Ebonite to determine structure of inputs and outputs for your model.
task = ebnt.get_or_create_task('my_project', 'my_task') model = task.create_and_push_model(clf, test_x, 'my_clf')
You are awesome! Now your model is safely persisted in a repository.
Later on in other Python process you can load your model from this repository and do some wonderful stuff with it, e.g., create a Docker image named my_service with an HTTP service wrapping your model.
from ebonite import Ebonite ebnt = Ebonite.local() task = ebnt.get_or_create_task('my_project', 'my_task') model = client.get_model('my_clf', task) client.build_image('my_service', model)
Check out examples (in examples directory) and documentation to learn more.
… is available here
… are available in this folder. Here are some of them:
Supported libraries and repositories
- your arbitrary Python function
- TensorFlow (1.x and 2.x)
- Model input / output data
- Model repositories
- local filesystem
- Amazon S3
Create an issue if you need support for something other than that!
Current release candidate
- Minor bugfixes
- Deleted accidental debug ‘print’ call :/
- Prebuilt flask server images for faster image build
- More and better methods in Ebonite client
- Pipelines - chain Models methods into one Model-like objects
- Refactioring of image and instance API
- Rework of pandas DatasetType: now with column types, even non-primitive (e.g. datetimes)
- Helper functions for stanalone docker build/run
- Minor bugfixes and features
- Fixed dependency inspection to include wrapper dependencies
- Fixed s3 repo to fail with subdirectories
- More flexible way to add parameters for instance running (e.g. docker run arguments)
- Added new type of Requirement to represent unix packages - for example, libgomp for xgboost
- Minor tweaks
- Minor fixes and examples update
- Built Docker images and running Docker containers along with their metadata are now persisted in metadata repository
- Added possibility to track running status of Docker container via Ebonite client
- Implemented support for pushing built images to remote Docker registry
- Improved testing of metadata repositories and Ebonite client and fixed discovered bugs in them
- Fixed bug with failed transactions not being rolled back
- Fixed bug with serialization of complex models some component of which could not be pickled
- Decomposed model IO from model wrappers
- bytes are now used for binary datasets instead of file-like objects
- Eliminated build_model_flask_docker in favor of Server-driven abstraction
- Sped up PickleModelIO by avoiding ModelAnalyzer calls for non-model objects
- Sped up Model.create by calling model methods with given input data just once
- Dataset types and model wrappers expose their runtime requirements
- Implemented asyncio-based server via aiohttp library
- Implemented support for Tensorflow 2.x models
- Changed default type of base python docker image to “slim”
- Added ‘description’ and ‘params’ fields to Model. ‘description’ is a text field and ‘params’ is a dict with arbitrary keys
- Fixed bug with building docker image with different python version that the Model was created with
- Fixed critical bug with wrapper_meta
- Fixed bug with deleting models from tasks
- Support working with model meta without requiring installation of all model dependencies
- Added region argument for s3 repository
- Support for delete_model in Ebonite client
- Support for force flag in delete_model which deletes model even if artifacts could not be deleted
- Eliminated tensorflow warnings. Added more tests for providers/loaders. Fixed bugs in multi-model provider/builder.
- Improved documentation
- Eliminate useless “which docker” check which fails on Windows hosts
- Perform redirect from / to Swagger API docs in Flask server
- Support for predict_proba method in ML model
- Do not fix first dimension size for numpy arrays and torch tensors
- Support for Pytorch JIT (TorchScript) models
- Bump tensorflow from 1.14.0 to 1.15.0
- Added more tests
- Multi-model interface bug fixes
- Minor bug fixes
- Added support for LightGBM models
- Added support for XGBoost models
- Added support for PyTorch models
- Added support for CatBoost models
- Added uwsgi server for flask containers
- Minor bug fixes
- First release on PyPI.
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Hashes for ebonite-0.7.0-py2.py3-none-any.whl