phobos 1.2.9

Training utility library and config manager for Granular Machine Vision research

A Utility Library that assists in Geospatial Machine Learning by:

• supporting creation of a project with boilerplate code for model training

• exporting annotations from Europa

• populating template project with configurable components for model

• fetching samples from dataset shards available at AIStore

• orchestrating model training and validation

• deploying project to Arche for efficient training in a node cluster

Flow

Features

• Polyaxon auto-param capture

• Configuration enforcement and management for translation into Dione environment

• Precomposed loss functions and metrics

• Get annotations from Europa

TODO

• ETL datasets via CLI on AIStore

• Multi Input and Multi Output models

• Static analysis code

• Dataset abstraction

• Standard dataset loaders

• Pretrained models

Build Details

• packages are managed using poetry

• packages poetry maintains pyproject.toml

• PRs and commits to develop branch trigger github actions

Tests

>>> make install
>>> make test-light

A GPU machine is requried for test-heavy

>>> make install
>>> make test-heavy

Installation

`pip install phobos`

Usage

Get all the annotation tasks available in Europa

`phobos get --all --email <email> --passwd <password>`

Download one particular annotation task from Europa

`phobos get --task <task ID> --path <directory to save anntoations> --email <email> --passwd <password>`

Create a project boilerplate code

`phobos init --project_name <project name> --project_description <project description>`

Run an experiment

`phobos run`

Run associated tensorboard

`phobos tensorboard --uuid <project id>`

License

GPLv3

Documentation

View documentation here

Image

Use gcr.io/granular-ai/phobos:latest