A framework for Automatic AI (AutoAI). Does automatic data pre-processing, model search and hyper-parameter turning. Comes with best in the industry code generation capability that produces high quality and well documented source code (train+test) for complete transparency and reproducible results.
Project description
BlobCity AutoAI
A framework to find, train and generate code for the best performing AI model. Works on Classification and Regression problems. The framework is currently designed for tabluar data, and is being extended to support images, videos and natural language.
Getting Started
pip install blobcity
If not already installed, then you must also install Tensorflow. This is a required depency, but not installed by default.
pip install tensorflow
import blobcity as bc
bc.train(file="data.csv", target="Y_column")
bc.spill("my_code.ipynb")
Y_column
is the name of the target column. The column must be present within the data provided.
Automatic inference of Regression / Classification is supported by the framework.
Supported input data formats are .csv
and .xlsx
. Extension for other file formats is being worked on.
The spill
function generates the model code with exhaustive documentation. Training code is also included for basic scikit-learn models. TensorFlow and other DNN models produce only the test / final use code.
Use a Pandas Data Frame
bc.train(df=my_df, target="Y_column")
If loading data from a Database or external system, create a DataFrame from your data source, and pass it directly to the train
function.
From a URL
model = bc.train(file="https://example.com/data.csv", target="Y_column")
The file
parameter can be a local file, or a URL. The function will load data from the specified URL. The file at the URL must be either in CSV or XLSX format. The URL should be accessible publicly without authentication.
Code Generation
Multiple formats of code generation is supported by the framework. The spill
function can be used to generate both ipynb
and py
files. The desired type is infered from the name of the output file. The code file will be created at the path specified. Relative and absolute file paths are both supported.
Generate Jupyter Notebook
model.spill("my_code.ipynb");
Generates an ipynb file with full markdown documentation containing code explanations.
Generate Python Code
model.spill("my_code.py")
Generates a Python code file. Code documentation is intentionally avoided by default to keep the Python code clean.
model.spill("my_code.py", docs=True)
Pass the optional docs
parameter to specify if relevant source code documentation should be included in the generated code.
Specifying Features
Framework automatically performs a feature selection. All features (except target) are considered by default for feature selection. Framework is smart enough to remove ID / Primary key columns.
You can manually specifiy a subset of features to be used for training.
bc.train("data.csv", target="Y_value", features=["col1", "col2", "col3"])
This does not guarantee that all specified features will be used in the final model. The framework will perform an automated feature selection from amongst these features. This only guarantees that other features if present in the data will not be considered.
Printing Model Stats
model.stats()
Key models parameters, such as Precision, Recall, F1-Score are printed using the stats
function. The parameters change based on the type of AutoAI problem.
Saving the Model
model.save('./my_model.pkl')
Use the save
method to serialise and save the model instance to a Pickle file. A trained model along with all its attributes can be saved. A saved model can be loaded back in the future, and used for preditions, code generation, or viewing the model stats.
Loading a Saved Model
model = bc.load('./my_model.pkl')
The loaded model is an exact replica of the model, as at the time of saving the model. The training state along with other training parameters are persisted and reloaded.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.