Agnostic model loader
Model Agnostic (MA) provides an easy model instantiation for Flask, Django, and other production Python projects. It also supports exporting models to other common serving platforms. Finally MA, allows you to easily load models and perform predictions without having to worry about the model backends. Moreover, it helps you organize your code for easy updating and maintainance.
Usage
To install run
pip install model_agnostic
You can also build from the source by (though we strongly reccomend using pip)
git clone https://github.com/isaacmg/model_agnostic_prediction.git
cd model_agnostic_prediction
python setup.py build install
To use create a new class that extends your model's backend subclass. Models are located in the model folder in the "frameworkName_serve" format.
from model_agnostic.models.pytorch_serve import PytorchModel
class NewPyTorch(PytorchModel):
def __init__(self, weight_path=None):
super(NewPyTorch, self).__init__(weight_path)
# Handle any model specific loading here
def preprocessing(self, raw_data):
# Apply necessary pre-processing here for your data here.
pass
def process_result(self):
# Implement this function if have any post-processing needs of predictions.
pass
You will have to implement preprocessing and possibly create_model and process_result depending on your model (see examples for more info).
Project Goals
-
Allow users to pre-load model weights from any framework with any save configuration
-
Works with common frameworks (PyTorch, Keras, Tensorflow, CNTK).
-
Same high level functions regardless of backend framework (i.e. preprocess, predict, process_response)
-
Provide a standard template which users can extend for their model's specific functionality.
-
Easy to use preprocessing functions that user can tweak depending on their model.
-
A model library (of filled out templates pre-setup) so user can deploy standard models easily onto Flask/Django.
-
Easy model exporting to ONNX, TF-Serving