Skip to content

Instantly share code, notes, and snippets.

@ZaxR
Last active January 13, 2020 19:49
Show Gist options
  • Save ZaxR/ad6eccfaeba3a98d2273200d9d9b5359 to your computer and use it in GitHub Desktop.
Save ZaxR/ad6eccfaeba3a98d2273200d9d9b5359 to your computer and use it in GitHub Desktop.
Attempt at a general ML model architecture
""" Demo Model class that can be used generally for ML projects.
Attempts to solve the following problems:
1. Ensure data preprocessing is consistent between data for model training and prediction
2. Have a common architecture for any data types / ML model types/libraries (as long as a consistent API is ued)
3. Allow for easy swapping of preprocessing, modeling, and/or postprocessing
4. Enforce an input (and potentially output) data schema(s)
Bonus:
- Follow's sklearn's estimator/transformer/predictor APIs, allowing use of sklearn Pipelines and GridSearchCV.
"""
from pyspins.io import load_file, output_file
from schema import Schema
from sklearn.base import BaseEstimator
# The below imports are for the demo only
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import MinMaxScaler, StandardScaler
class NoModelError(Exception):
pass
class MissingFilePathError(Exception):
pass
class MLModelError(Exception):
""" Exception type used to raise exceptions within MLModel derived classes """
def __init__(self, *args):
Exception.__init__(self, *args)
class MLModelSchemaValidationError(MLModelError):
""" Exception type used to raise schema validation exceptions within MLModel derived classes """
def __init__(self, *args):
MLModelError.__init__(self, *args)
class Model(BaseEstimator): # maybe don't inherit from BaseEstimator to remove sklearn dependency
"""Build, run predictions on, and store metadata for a ML model.
Args:
model: ML model with a `predict` method.
input_schema: See https://pypi.org/project/schema/
preprocessor: One or more preprocessing steps for fitting/predicting data.
Must have a fit_transform method.
Works with sklearn Pipelines and Transformers.
E.g. `preprocessor=Pipeline([('minmax', MinMaxScaler()),
('std', StandardScaler())])`
trainer: One or more model ML estimators.
Must have a fit method.
Works with sklearn Pipelines and Estimators.
E.g. `trainer=RandomForestClassifier()`
postprocessor: One or more postprocessing steps for transforming predictions.
Must have a fit_transform method.
Works with sklearn Pipelines and Transformers.
model_path:
"""
def __init__(self, model=None, input_schema=Schema(None),
preprocessor=None, trainer=None, postprocessor=None, model_path=None):
self.model = model
self.input_schema = input_schema
# preprocessing may depend on self.training; postprocess will also depend on self.training
self.preprocessor = preprocessor
self.trainer = trainer
self.postprocessor = postprocessor
self.model_path = model_path
def fit(self, X, y, **train_kwargs):
"""
Note:
Can be used either to train a single model or with GridSearchCV.
"""
X = self.preprocessor.fit_transform(X)
model = self.trainer.fit(X, y, **train_kwargs)
self.model = model
# return self?
def predict(self, prediction_inputs):
try:
self.input_schema.validate(prediction_inputs)
except Exception as e:
raise MLModelSchemaValidationError("Failed to validate input data: {}".format(str(e)))
if self.preprocessor is not None:
prediction_inputs = self.preprocessor.fit_transform(prediction_inputs)
predictions = self.model.predict(prediction_inputs)
if self.postprocessor is not None:
predictions = self.postprocessor.transform(predictions)
return predictions
def save(self):
if self.model is None:
raise NoModelError
if self.model_path is None:
raise MissingFilePathError
output_file(self, self.model_path)
@classmethod
def load(cls, model_path):
model_instance = load_file(model_path)
if isinstance(model_instance, cls):
return model_instance
raise TypeError("The file at model_path isn't an instance of Model.")
# Demo
if __name__ == '__main__':
df = pd.DataFrame({"feature1": [1, 2, 3, 4, 5, 6, 7, 8],
"target": [1, 1, 1, 1, 2, 2, 2, 2]})
target = ["target"]
model_path = "modeltestv1.pkl"
# Training + Predicting
X_train, X_test, y_train, y_test = train_test_split(df.drop(target, axis="columns"),
df[target],
test_size=0.2)
preprocessor = Pipeline([('minmax', MinMaxScaler()), ('std', StandardScaler())])
trainer = RandomForestClassifier()
# postprocessor = PredCleaner
model = Model(model=None,
input_schema=Schema(pd.DataFrame),
preprocessor=preprocessor,
trainer=trainer,
postprocessor=None,
model_path=model_path)
model.fit(X_train, y_train)
model.save()
preds = model.predict(X_test)
actuals = y_test
# Predicting from existing model
to_predict = pd.DataFrame([100, 0])
model = Model.load(model_path)
preds = model.predict(to_predict)
@ZaxR
Copy link
Author

ZaxR commented Jan 13, 2020

Tensorflow - issues with saving model. It wants to save lots of files, not a single serialized object

@ZaxR
Copy link
Author

ZaxR commented Jan 13, 2020

2 votes for setting path in the method, not the class insitantiation

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment