sapsan 0.0.2a0

Sapsan project

Sapsan

• Intro
• Structure
  • Estimator
  • Experiment
  • Dataset
  • Tracking backend
• Examples
• Kubeflow

---

Intro

Sapsan is a pipeline for easy Machine Learning implementation in scientific projects. That being said, its primary goal and featured models are geared towards dynamic MHD turbulence subgrid modeling. Sapsan will soon feature Physics-Informed Machine Learning models in its set of tools to accurately capture the turbulent nature appicable to Core-Collapse Supernovae.

Note: currently Sapsan is in alpha, but we are actively working on it and introduce new feature on a daily basis.

Getting Started

To get started, simply pull sapsan via pip (or you could clone this repo)

pip install sapsan

Next, please run an example to make sure everything has been installed correctly. It is a jupyter notebook which can be found here:

Sapsan/examples/cnn_example.ipynb

In order to get started on your own project, you can use the command-line-interface interface:

sapsan create --name awesome

---

Structure

Structure of project is build around few concepts making this project easier to extend to more cases. Core abstractions are: estimator, dataset, experiment, tracking backend Core abstraction are defined in models.py file.

Estimator

General abstraction for models/algorithms.

Available estimators

• KRR 1d estimator
• 3d convolution encoder estimator
• 3d autoencoder estimator (coming soon!)

How to implement new estimator:

Extend Estimator class and implement train, predict and metrics methods.

from sapsan.core.models import Estimator, EstimatorConfiguration


class AwesomeEstimator(Estimator):
    def __init__(self, config: EstimatorConfiguration):
        super().__init__(config)
        self.model = ... # your awesome model

    def train(self, inputs, labels):
        # fit model to labels
        return self.model

    def predict(self, inputs):
        prediction = ... # derive prediction from inputs
        return prediction

    def metrics(self):
        return {"training_time": 146, "training_avg_loss": 42}

Dataset

General abstraction for dataset/dataframes.

Available datasets

• 3d dataset
• 2d dataset (coming soon!)

How to implement new dataset:

Extend Dataset class and impement load method.

import numpy as np
from sapsan.core.models import Dataset


class RandomDataset(Dataset):
    def __init__(self, n_entries: int, n_features: int):
        self.n_entries = n_entries
        self.n_features = n_features

    def load(self):
        return np.random.random((self.n_entries, self.n_features))

Experiment

General abstraction for experiments.

Available experiments

• general training experiment
• evaluation 1d experiment
• evaluation 3d encoder experiment
• evaluation 3d autoencoder experiment

How to implement new experiment:

Extend Experiment class and impement run method.

from sapsan.core.models import Experiment, ExperimentBackend


class AwesomeExperiment(Experiment):
    def __init__(self, name: str, backend: ExperimentBackend):
        super().__init__(name, backend)

    def run(self):
        # do whatever you need to execute during experiment
        return {}

Tracking backend

General abstraction for experiment tracker.

Available tracking backends

• MlFlow
• FakeBackend

How to implement new experiment:

Extend ExperimentBackend class and impement log_metric, log_parameter, log_artifact methods.

from sapsan.core.models import ExperimentBackend


class InMemoryTrackingBackend(ExperimentBackend):
    def __init__(self, name: str):
        super().__init__(name)
        self.metrics = []
        self.parameters = []
        self.artifacts = []

    def log_metric(self, name, value):
        self.metrics.append((name, value))

    def log_parameter(self, name, value):
        self.metrics.append((name, value))

    def log_argifact(self, path):
        self.artifacts.append(path)

---

Examples

Examples of implemented experiments.

• 3d convolution encoder example
• 3d audoencoder example (coming soon!)
• KRR 1d estimator (coming soon!)

---

Kubeflow

Docs for experiments in kubeflow

Examples

Local via docker compose

docker-compose build
docker-compose up --force-recreate

Then open browser at localhost:8888

---

© (or copyright) 2019. Triad National Security, LLC. All rights reserved. This program was produced under U.S. Government contract 89233218CNA000001 for Los Alamos National Laboratory (LANL), which is operated by Triad National Security, LLC for the U.S. Department of Energy/National Nuclear Security Administration. All rights in the program are reserved by Triad National Security, LLC, and the U.S. Department of Energy/National Nuclear Security Administration. The Government is granted for itself and others acting on its behalf a nonexclusive, paid-up, irrevocable worldwide license in this material to reproduce, prepare derivative works, distribute copies to the public, perform publicly and display publicly, and to permit others to do so.