Welcome to Cascade!#

Lightweight and modular MLOps library with the aim to make ML development more efficient targeted at small teams or individuals.

Who could find it useful#

The slope of adopting MLOps platforms can sometimes be too steep for small teams. However, they can still benefit from MLOps practices integrated into the workflow. Cascade offers the solution that enables those features for small projects while demanding little.

Key Principles#

  • Elegancy - ML code should be about ML with minimum meta-code

  • Agility - it should be easy to build prototypes and integrate existing project with Cascade

  • Reusability - code should have an ability to be reused in similar projects

  • Traceability - everything should have meta data

Introduction#

Cascade does not require any complex setups. To start use it you just need to install the python package.

Installation#

pip install cascade-ml

More info on installation can be found in Quickstart page.

Track your first experiment#

Cascade is built to be modular and flexible so that integration of Cascade in your existing workflow could become gradual.

The simplest case is the experiment tracking - store your models structured along with all meta data.

from cascade.models import BasicModel, ModelRepo
import random


model = BasicModel()
model.params.update({"learning_rate": 1e-10})

# Your existing experiment goes here

model.link(name="dataset", uri="../data/dataset")
model.add_metric("f1", random.random())
model.add_file("example_plot.png")

repo = ModelRepo("classification")
line = repo.add_line("resnet")
line.save(model)

This brief example shows that you can easily start experiment tracking from your current setup.

Cascade allows you to abstract from structuring your model storage. It creates and manages repositories of models from which they can be extracted easily for evaluation or deployment.

|- classification
|  |- resnet
|  |  |- 00000
|  |  |  |- artifacts
|  |  |  |- files

What is saved as default meta data in classification/resnet/00000/meta.json 

[
    {
        "comments": [],
        "created_at": "2023-09-18T19:43:48.146504+00:00",
        "description": null,
        "links": ["name": "dataset", "uri": "../data/dataset", "meta": null],
        "metrics": [{"created_at": "2023-09-18T19:44:39.853785+00:00",
                    "dataset": null,
                    "direction": null,
                    "interval": null,
                    "name": "f1",
                    "split": null,
                    "value": 0.45756550190760725}],
        "name": "cascade.models.model.Model",
        "params": {"learning_rate": 1e-10},
        "path": "/home/ilia/classification/resnet/00000",
        "saved_at": "2023-09-18T19:46:00.099433+00:00",
        "slug": "perky_awesome_coua",
        "tags": [],
        "type": "model"
    }
]

Meta data is a very flexible structure and you can save and modify whatever information you want.

For more advanced usage of Model see Model training page

Build pipelines#

The amount of work we put in the building data processing pipelines is proportional to the role clean and well prepared data plays in the final performance of ML solution.

This is where Datasets come in. Consider this toy-example where the data would be a sequence of digits.

Cascade offers a library of useful default datasets and modifiers along with a modular paradigm of independent pipeline steps as modifiers.

from cascade import data as cdd


ds = cdd.Wrapper([0, 1, 2, 3, 4])
ds = cdd.ApplyModifier(ds, lambda x: x**2)

In this example our data goes through three stages that are defined above. The pipeline is defined in a declarative way and each step will be executed as the particular item requested.

Though the main feature of datasets is their metadata. To get metadata of an object you do just:

train_ds.get_meta()

[
    {
        "comments": [],
        "description": null,
        "len": 5,
        "links": [],
        "name": "cascade.data.apply_modifier.ApplyModifier",
        "tags": [],
        "type": "dataset"
    },
    {
        "comments": [],
        "description": null,
        "len": 5,
        "links": [],
        "name": "cascade.data.dataset.Wrapper",
        "obj_type": "<class 'list'>",
        "tags": [],
        "type": "dataset"
    }
]

By default all sized objects will have length in their meta data. The more specific the transformation about the dataset the richer meta will be.

All datasets with examples can be found in Dataset Zoo