auto-ml-cl 0.0.9

Auto machine learning with scikit-learn and TensorFlow framework.

auto-ml-cl

How to create a machine learning and deep learning models with just a few lines of code by just provide data, then framework will get best trained models based on the data we have? We don't need to care about Data Loading, Feature Engineering, Model Training, Model Selection, Model Evaluation and Model Sink, even RESTful with best trained model. Now this is auto_ml_cl comes in.

This repository is based on scikit-learn and TensorFlow to create both machine learning models and nueral network models with few lines of code by just providing a training file, if there is a test file will be nicer to evaluate trained model without any bias, but if with just one file will also be fine. Currently this repository is only support with Classification problem.

Key features highlights:

• machine learning and neural network models are supported.
• Automatically data pre-processing with missing, unstable, categorical various data types.
• Ensemble logic to combine models to build more powerful models.
• Nueral network models search with kerastunner to find best hyper-parameter for specific type of algorithm.
• Cloud files are supported like: Cloud storage for GCP or local files.
• Logging different processing information into one date file for future reference.
• Processing monitoring for each algorithm training status.
• RESTful API for API call to get prediction based on best trained model.

Sample code to use auto_ml package by using Titanic dataset from Kaggle competion, as this dataset contain different kinds of data types also contain some missing values with different threasholds.

from auto_ml.automl import ClassificationAutoML, FileLoad

file_name = 'train.csv'
file_path = r"C:\auto_ml\test"  # Absolute path
file_load = FileLoad(file_name, file_path=file_path, label_name='Survived')

auto_cl = ClassificationAutoML()
auto_cl.fit(file_load=file_load, val_split=0.2)

# Get prediction based on best trained models
test_file_name = 'test.csv'
file_load_test = FileLoad(test_file_name, file_path=file_path)

pred = auto_cl.predict(file_load=file_load_test)

Then we could get whole trained models' evaluation score for each trained model score, we could get best trained model based on validation score if we would love to use trained model for production, one important thing is that these models are stored in local server, we could use them any time with RESTFul API calls.

If we want to use GCP cloud storage as a data source for train and test data, what needed is just get the service account file with proper authority, last is just provide with parameter: service_account_name and file local path: service_account_file_path to FileLoad object, then training will start automatically.

file_name="train.csv"
file_path = "gs://bucket_name"
service_account_name = "service_account.json"
service_account_file_path = r"C:\auto_ml\test"

file_load = FileLoad(file_name, file_path, label_name='Survived', 
    service_account_file_name=service_account_name, service_account_file_path=service_account_file_path)

auto_cl = ClassificationAutoML()
auto_cl.fit(file_load=file_load)

If we have data in memory, we could also use memory objects to train, test and predict with auto_ml object, just like scikit-learn.

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split

x, y = load_iris(return_X_y=True)
xtrain, xtest, ytrain, ytest = train_test_split(x, y, test_size=.2)

auto_cl = ClassificationAutoML()
auto_cl.fit(xtrain, ytrain)

score = auto_cl.score(xtest, ytest)
pred = auto_cl.predict(xtest)
prob = auto_cl.predict_proba(xtest)

Current supported algorithms:

• Logistic Regression
• Support vector machine
• Gradient boosting tree
• Random forest
• Decision Tree
• Adaboost Tree
• K-neighbors
• XGBoost
• LightGBM
• Deep nueral network

Also supported with Ensemble logic to combine different models to build more powerful model by adding model diversity:

• Voting
• Stacking

For raw data file, will try with some common pre-procesing steps to create dataset for algorithms, currently some pre-processing algorithms are supported:

• Imputation with statistic analysis for continuous and categorical columns, also support with KNN imputaion for categorical columns.
• Standarize with data standard data
• Normalize
• OneHot Encoding for categorical columns
• MinMax for continuous columns to avoid data volumn bias
• PCA to demension reduction with threashold
• Feature selection with variance or LinearRegression or ExtraTree

Insight for logics to auto machine learning training steps.

1. Load data from file or memory for both training and testinig with class FileLoad, support with GCP's GCS files as source file.

2. Build processing pipeline object based on data.

   (1). Imputation for both categorical and numerical data with different logic, if data missing column is over a threshold, will delete that column. Support with algorithm KNNImputer to impute data or SimpleImputer to fill missing data.

   (2). OneHot Encoding for categorical columns and add created columns into original data.

   (3). Standardize data to avoid data range, also benefit for some algorithms like SVM etc.

   (4). MinMax data to keep data into a 0-1 range.

   (5). FeatureSelection to keep features with a default threshold or using algorithm with ExtraTree or LinearRegreesion to select features.

   (6). PCA to reduce dimenssion if feature variance over a threshold and just keep satisfied features.

3. Build a Singleton backend object to do file or data related functions.

4. Build training pipeline to instant each algorithm with a factory class based on pre-defined used algorithms.

5. Build a SearchModel class for each algorithm to find best parameters based on RandomSearch or GridSearch.

6. Pre-processing pipeline fit and tranform, save trained pipeline into disk for future use.

7. Start training with training pipeline with processed data with doing parameters search to find best parameter's model, also combined with Neural network search to find best neural models. If need validation will use some data to do validation that will reduce training data size, or could use trainded auto_ml object to do validation will also be fine.

8. Use Ensemble logic to do voting or stacking to combine trained models as a new more diverse model based on best trained model.

9. Evaluate each trained models based on validation data and return a ditionary with training model name, training score and validation score.

10. Support to export trained models into a new folder that we want.