Docker template for evaluating Tensorflow.Keras ML Models

The purpose of this tool is to provide an OpenSource option to evaluate Tensorflow models without Tensorflow installation issues.

• Local requirements:
  • Docker installed (see Docker)
  • docker-compose installed (see docker-compose)
• Features:
  • Custom packages can be installed in python image.
  • Download weights and/or model from Google Drive.
  • Load model from pickle, native keras formats or tensorflow.
  • Evaluate your model.

Getting Started

1. First define the library requirements for loading and evaluating your model in requierements.txt. (Just as you would install with pip3 install <library>)
2. Run docker-compose build to create the image with all the requierements and environment for your evaluation.
3. Add your model code to model.py and import it in evaluate.py.
4. Upload weights or model to Googe Drive storage or locally if you prefer. (.pkl, .h5 or SavedModel supported)
5. Update config.json according to your files.
6. Run evaluation with docker-compose up web

Example config.json

---

Create empty model and load weights from .pkl

{
    "drive":[
            { "file_id":"FILE_ID_HERE", "destination":"model_weights.pkl"}
          ],
    "model":{
        "required":false,
        "location":"",
        "pickle":false
    },
    "weights":{
        "required":true,
        "location":"model_weights.pkl",
        "pickle":true
    }     
}

References:

• https://stackoverflow.com/a/39225039