A Docker image for serving TensorFlow models via a REST API.

TensorFlow Serving enables you to deploy and serve TensorFlow models via a REST API, making it easier to make predictions using your trained models.

For a comprehensive understanding of TensorFlow Serving, you can watch this informative YouTube video and refer to this GitHub repository.

To use this TensorFlow Serving Docker image, follow these steps:

1. Clone this repository as a Git submodule in your project or clone it separately.

2. Organize your models in the following folder structure:

   saved-models/
   ├── models/               # This is where TensorFlow saves your models
   └── tf-serving-configs/   # Save your TensorFlow Serving configurations (sample configurations in [config-samples/](config-samples))
   

3. Create a local .env file inside this repository with the necessary environment variables:

   # .env
   
   # Define the port for the REST API
   REST_API_PORT=8605
   
   # Model location in the host machine
   MODELS_LOCATION=../training/saved-models
   
   # The configuration file of the models to be used when loading the TensorFlow Serving API
   CONFIGURATION_FILE=models.config.a

   You can use the example from the sample.env file and customize it as needed.

Once you have set up the environment, you can start your api with:

$ docker compose up

Then you can interact with your TensorFlow Serving API using the following endpoints:

• Access model information: http://localhost:8605/v1/models/my-model (GET request)

To make predictions, you can use one of the following URLs:

• Predict by model version number: http://localhost:8605/v1/models/my-model/versions/2:predict (POST request)
• Predict by model name: http://localhost:8605/v1/models/my-model/labels/beta:predict (POST request)

These endpoints enable efficient utilization of your TensorFlow models for prediction through the REST API.

When you want the model to make a prediction, you need to send a POST request to one of the above endpoints. The payload should adhere to the following format:

{
   "instances": [
      **instance_1**,
      **instance_2**,
      ...
   ]
}

The response to the POST request will be structured as follows:

{
   "predictions": [
      **prediction_1**,
      **prediction_1**,
      ...
   ]
}