sclblpy 0.1.7

Python package for uploading sklearn and onnx models to Scailable toolchain.

sclblpy

sclblpy is the core python package provided by Scailable to convert models fit in python to WebAssembly and open them up as a REST endpoint.

Currently the package supports the upload of fitted sklearn model objects, and it supports uploading .onnx models (by specifying a path to the .onnx file). The package can also be used to manage assignments of models to registered devices.

sclblpy is only functional in combination with a valid Scailable user account.

• Website: https://www.scailable.net
• Docs:
  • On github: https://github.com/scailable/sclblpy
  • On pypi: https://docs.sclbl.net/sclblpy
  • API docs Scailable: https://docs.sclbl.net
• Get an account: https://www.scailable.net
• Source: https://github.com/scailable/sclblpy/
• Getting started: https://github.com/scailable/sclbl-tutorials/tree/master/sclbl-101-getting-started

Background

The sclblpy package allows users with a valid scailable account (apply for one at https://www.scailable.net) to upload fitted ML / AI models to the Scailable toolchain server. This will result in:

1. The model being tested on the client side.
2. The model being uploaded to Scailable, tested again, and if all test pass it will be converted to WebAssembly.
3. The model being made available as an easy to access REST endpoint.

Getting started

The following functions are likely most used:

sp.upload() can be used to create model:

def upload(mod, features, docs={}, email=True, model_type="sklearn", _keep=False) -> bool:
    """upload uploads a trained AI/ML model to Scailable.

    The upload function is the main workhorse of the sclblpy package but effectively provides a
    wrapper to choose between the
     - upload_sklearn(mod, feature_vector, docs={}, email=True, _keep=False)
     - upload_onnx(path, docs={}, email=True)
    functions.

    The function checks the type, and if type = "sklearn" (default) calls the upload_sklearn() function.
    If type = "onnx" it calls the upload_onnx() function.

    Args:
        mod: The model to be uploaded (type="sklearn" OR the path to the stored ONNX file (type="onnx").
        features:
            - An example feature_vector for your model (type="sklearn" only).
            (i.e., the first row of your training data X obtained using row = X[0,:])
            - The input str (binary) to the onnx model (type="onnx" only). Can be an empty string.
        docs: A dict{} containing the fields 'name' and 'documentation'.
        email: Bool indicating whether a confirmation email of a successful conversion should be send. Default True.
        model_type: String indicating the type of model. Currently with options "sklearn" or "onnx". Default "sklearn"
        _keep: Bool indicating whether the .gzipped file should be retained (type="sklearn" only). Default False.

    Returns:
        False if upload failed, true otherwise"""

sp.models() lists all created models, and sp.delete_model() can be used to delete a model. Finally, sp.update() can be used to overwrite / update an existing model.

sp.assign() can be used to assign a model to a device:

def assign(cfid, did, rid, _verbose=True):
    """ Assign a model to a device.

    Using the global JWT string this function assigns a model (using its cfid) to a device (using its did).

    Args:
        cfid: String identifying the model / compute-function
        did: String identifying the device
        rid: String identifying the registration ID of the device (not, run "devices"

    Returns:
        Boolean indicating whether the assignment was successful.
    """

sp.assignments() can be used to list current assignments, whereas sp.delete_assignment() can be used to delete an assignment.

A simple example using sklearn

After installing the package using pip install sclblpy you can easily fit a ML / AI model using your preferred tools and upload it to our toolchains. The following code block provides a simple example:

# Neccesary imports:
import sclblpy as sp

from sklearn import svm
from sklearn import datasets

# Start fitting a simple model:
clf = svm.SVC()
X, y = datasets.load_iris(return_X_y=True)
clf.fit(X, y)

# Create an example feature vector (required for sklearn models):
row = X[130, :]

# Create documentation (optional, but useful):
docs = {}
docs['name'] = "My first fitted model"
docs['documentation'] = "Any documentation you would like to provide."

# Upload the model:
sp.upload(clf, row, docs=docs)

The call to sp.upload() will upload the fitted model, after running a number of local tests, to the Scailable toolchain server and create an associated REST endpoint. Limited user feedback will be printed to show progress, and you will receive an email at the email address associated with your account when the conversion is fully completed (which might take a few minutes). This email also contains further details regarding the usage of your created endpoint.

Note that upon first upload you will be prompted to provide your Scailable username and password; you can choose to store the provided credentials locally to enable easy login on subsequently uploads. (users can signup for an account at https://www.scailable.net).

More sklearn examples

These examples are merely intended to show the desired syntax for the various packages; we do not intend to fit models that actually have a good predictive performance in these examples.

Currently we support uploading sklearn, statsmodels, and xgboost models (run sp.list_models() to print an overview of all supported models). Here we provide an example for each of these.

sklearn: the elastic net

The elastic net provides a flexible regularized model that is useful for many supervised learning tasks.

import sclblpy as sp
from sklearn import linear_model  # Import linear_model
from sklearn import datasets  # Import sklearn datasets

iris_data = datasets.load_iris(return_X_y=True)
X, y = iris_data

mod = linear_model.ElasticNet()  # Instantiate the model

mod.fit(X, y)  # Fit the model

fv = X[0, :]  # An example feature vector
docs = {'name': "ElasticNet example model", 'documentation' : "Documentation for this model."}

sp.upload(mod, fv, docs)

statsmodels: OLS regression

The statsmodels package provides a number of regression models with flexible error functions. We provide a simple OLS example:

import sclblpy as sp

import statsmodels.api as sm  # Import statsmodels
from sklearn import datasets  # Import sklearn datasets

iris_data = datasets.load_iris(return_X_y=True)
X, y = iris_data

est = sm.OLS(y, X)  # Specify the model
mod = est.fit()  # Fit the model; note that we send the fitted model to scailable

fv = X[0, :]  # An example feature vector
docs = {'name': "OLS example model"}

sp.upload(mod, fv, docs=docs)

xgboost: tree boosting

The xgboost package provides flexibly tree boosting models (see xgboost). This model often performs very well "off the shelf" for many supervised learning tasks.

import sclblpy as sp

from xgboost import XGBClassifier  # Import xgboost classifier
from sklearn import datasets  # Import sklearn datasets

iris_data = datasets.load_iris(return_X_y=True)
X, y = iris_data

mod = XGBClassifier()  # Instantiate the model
mod.fit(X, y)  # Fit the model

fv = X[0, :]  # An example feature vector
docs = {'name': "XGBoost example model"}

sp.upload(mod, fv, docs=docs)

A simple .onnx example

The following code can be used to upload a stored .onnx model:

# Add docs
docs = {}
docs['name'] = "Name of ONNX model"
docs['documentation'] = "A long .md thing...."
check = sp.upload("PATH-TO-MODEL/FILE-NAME.onnx", "", docs, model_type="onnx")

Note that the file will be send to the Scailable platform; after it has been transpiled to WebAssembly you will receive (by default) and email.

Additional functionality

Next to the main upload() function, the package also exposes the following functions to administer endpoints:

# List all models owned by the current user:
sp.models()

# Remove an endpoint:
sp.delete_models(cfid)  # Where cfid is the compute function id

# Update an existing endpoint:
sp.update(mod, fv, cfid, docs)  # Where cfid is the compute function id

# Update an existing ednpoint without updating the docs:
sp.update(mod, fv, cfid) 

# Update only the docs of an existing endpoint:
sp.update_docs(cfid, docs)

# See all devices:
sp.devices(offset=0, limit=20, _verbose=True, _return=False)

# Delete device:
sp.delete_device(did)

# See all assignments:
sp.assignments(offset=0, limit=20, _verbose=True, _return=False)

# Create an assignment:
sp.assign(cfid, did, rid, _verbose=True)

# Remove an assignment:
sp.delete_assignment(aid) 

Additionally, the following methods are available:

# List all models currently supported by our toolchains:
sp.list_models()  

# Prevent any user feedback from being printed:
sp.stop_print()  

# Turn user feedback back on:
sp.start_print()  

# Remove locally stored user credentials:
sp.remove_credentials()

Running an uploaded model

After uploading a model to Scailable using the sclblpy package, you might also want to use python to consume the model. You will find example python code to consume your created endpiont in your Scailable admin which you can directly copy-paste into your python project. But, if you want it even easier you can also add the following to your code:

from sclblpy import run

cfid = "e93d0176-90f8-11ea-b602-9600004e79cc"  # This is the integer sum demo.
fv = [1,2,3,4,5]
result = run(cfid, fv)

print(result) # Prints the full result from the Scailable server.

Supported sklearn models

The list of models supported by the current version of the sclblpy package can always be retrieved using the list_models() function. Here we provide an overview:

Package	Model	Tested (19-06-2020)	Note
lightgbm	LGBMClassifier	ok
lightgbm	LGBMRegressor	ok
sklearn	ARDRegression	ok
sklearn	BayesianRidge	ok
sklearn	DecisionTreeClassifier	ok
sklearn	DecisionTreeRegressor	ok
sklearn	ElasticNet	ok
sklearn	ElasticNetCV	ok
sklearn	ExtraTreeClassifier	ok
sklearn	ExtraTreeRegresso	ok
sklearn	ExtraTreesClassifier	ok
sklearn	ExtraTreesRegressor	ok
sklearn	HuberRegressor	ok
sklearn	Lars	ok
sklearn	LarsCV	ok
sklearn	Lasso	ok
sklearn	LassoCV	ok
sklearn	LassoLars	ok
sklearn	LassoLarsCV	ok
sklearn	LassoLarsIC	ok
sklearn	LinearRegression	ok
sklearn	LinearSVC	ok
sklearn	LinearSVR	ok
sklearn	LogisticRegression	ok
sklearn	LogisticRegressionCV	ok
sklearn	NuSVC	ok
sklearn	NuSVR	ok
sklearn	OrthogonalMatchingPursuit	ok
sklearn	OrthogonalMatchingPursuitCV	ok
sklearn	PassiveAggressiveClassifier	ok
sklearn	PassiveAggressiveRegressor	ok
sklearn	Perceptron	ok
sklearn	RandomForestClassifier	ok
sklearn	RandomForestRegressor	ok
sklearn	RANSACRegressor	ok
sklearn	Ridge	ok
sklearn	RidgeClassifier	ok
sklearn	RidgeClassifierCV	ok
sklearn	RidgeCV	ok
sklearn	SGDClassifier	ok
sklearn	SGDRegressor	ok
sklearn	SVC	ok
sklearn	SVR	ok
sklearn	TheilSenRegressor	ok
statsmodels	Generalized Least Squares (GLS)	ok
statsmodels	Generalized Least Squares with AR Errors (GLSAR)	ok
statsmodels	Ordinary Least Squares (OLS)	ok
statsmodels	Quantile Regression (QuantReg)	ok
statsmodels	Weighted Least Squares (WLS)	ok
xgboost	XGBClassifier	ok
xgboost	XGBRegressor	ok
xgboost	XGBRFClassifier	ok	Binary only
xgboost	XGBRFRegressor	ok

Supported .onnx files

We currently support ONNX version 1.8 (and below) in full. However, if you encounter any problems converting .onnx files please let us know.

Dependencies

sclblpy needs python 3, and has been tested on python > 3.7. Furthermore, dependent on usage, sclblpy will import the following packages:

• numpy
• requests
• uuid
• sklearn

The statsmodels and xgboost packages are imported when used. No onnx packages are neccesary for the sclblpy package to run.

Notes:

• We try to stick to the naming conventions in http://google.github.io/styleguide/pyguide.html.
• The methods _set_toolchain_URL(string) and _set_usermanager_URL(string) can be used to change the default location of the toolchain and user-management function. These are useful when running the Scailable stack locally. Also the method _toggle_debug_mode() can be used for troubleshooting (this will raise exceptions and provide a trace upon errors).
• Docs generated using pdoc --html --html-dir docs sclblpy/main.py

If you are having trouble using the sclblpy package, please submit an issue to our github, we will try to fix it as quickly as possible!