Build Low Code Automated Tensorflow, What-IF explainable models in just 3 lines of code.
We aim to make Deep Learning on Tensorflow absolutely easy for the masses with our low code framework and also increase trust on ML models through What-IF tool simulation.
Built on top of the powerful Tensorflow ecosystem like TFX , TF APIs and What-IF Tool , the library automatically does all the heavy lifting internally like EDA, schema discovery, HPT, model search etc. This empowers developers to focus only on building end user applications quickly without any knowledge of Tensorflow, ML or debugging. There is no dependency on Pandas / SKLearn or other libraries which makes the whole pipeline highly scalable on any volume of data. Moreover the models trained with auto-tensorflow can directly be deployed on any cloud like GCP / AWS / Azure.
- Build Classification / Regression models on CSV data
- Automated Schema Inference
- Automated EDA and visualization
- Automated Model build for mixed data types( Continuous, Categorical and Free Text )
- Automated Hyper-parameter tuning
- Automated UI based What-IF analysis
- Control over complexity of model
- No dependency over Pandas / SKLearn
- Can handle dataset of any size - including multiple CSV files
- Install library using -
!pip install auto-tensorflow
- Works best on UNIX/Linux/Debian/Google Colab
- Initialize TFAuto Engine
from auto_tensorflow.tfa import TFAuto
tfa = TFAuto(train_data_path='/content/train_data/', test_data_path='/content/test_data/', path_root='/content/tfauto')
- Step 1 - Automated EDA and Schema discovery
tfa.step_data_explore(viz=True) ##Viz=False for no visualization
- Step 2 - Automated ML model build and train
tfa.step_model_build(label_column = 'price', model_type='REGRESSION', model_complexity=1)
- Step 3 - Automated What-IF Tool launch
tfa.step_model_whatif()
Tutorials: https://github.com/rafiqhasan/auto-tensorflow/tree/main/tutorials
-
Method TFAuto
train_data_path
: Path where training data is storedtest_data_path
: Path where Test / Eval data is storedpath_root
: Directory for running TFAuto( Directory should NOT exist )
-
Method step_data_explore
viz
: Is data visualization required ? - True or False( Default )
-
Method step_model_build
label_column
: The feature to be used as Labelmodel_type
: Either of 'REGRESSION'( Default ), 'CLASSIFICATION'model_complexity
: 0 to 1 (0: Model without HPT, 1(Default): Model with HPT) -> More will be added in future
There are a few limitations in the initial release but we are working day and night to resolve these and add them as future features.
- Doesn't support Image / Audio data
- Doesn't support - quote delimited CSVs( TFX doesn't support qCSV yet )
- Classification only supports integer labels from 0 to N
- Add support for Timeseries / Audio / Image data
- Add support for quoted CSVs
- Add feature to download full pipeline model Python code for advanced tweaking