TensorFlow CPU

Name	tensorflow-cpu
Version	v1.0.0
DockerHub	weevenetwork/tensorflow-cpu
Authors	Jakub Grzelak

• TensorFlow CPU
  • Description
  • Environment Variables
    • Module Specific
    • Set by the weeve Agent on the edge-node
  • Dependencies
  • Input
  • Output

Description

Use your pre-trained TensorFlow (Keras or non Keras) model with weeve modules. This module supports only CPU models and we plan to make a separate module to run TensorFlow with CUDA. The model should be available to the module via a downloadable URL (zipped model folder) or it should be stored in the edge device local filesystem (unzipped or regular folder). The module will take input data and then compose a tensor with data in the order assigned to the Ordered Labels environment variable. Later, the module will input that tensor into the model.

If the model is Keras, then the module uses high-level API to load the model. Otherwise, the module uses low-level API (SavedModel).

The module uses a function provided in Prediction Function config (env FORWARD_PROP_FUNCTION) to make a prediction. For instance, if Prediction Function is predict then the module uses model.predict(X) to make a prediction, if Prediction function is __call__ then model(X), etc.

Environment Variables

Module Specific

The following module configurations can be provided in a data service designer section on weeve platform:

Name	Environment Variables	type	Description
Model ZIP Download URL	MODEL_ZIP_DOWNLOAD_URL	string	If model is stored online, then provide a download URL to parse the model in a zip file. Leave empty field to search for the model in the local filesystem.
Model Folderpath	MODEL_LOCAL_FOLDERPATH	string	If model is stored in the local filesystem (above field for URL was left empty), then provide a path to the folder containing the model.
Keras	IS_KERAS	string	Whether the model is Keras or not.
Prediction Function	FORWARD_PROP_FUNCTION	string	Name of the prediction (forward propagation function), i.e. predict or __call__
Input Labels	INPUT_LABELS	string	Input data labels in the order of feeding into the model. Later a TensorFlow tensor will be created to feed that data into the model in the given order.
Output Label	OUTPUT_LABEL	string	The output label at which data is dispatched.

Set by the weeve Agent on the edge-node

Other features required for establishing the inter-container communication between modules in a data service are set by weeve agent.

Environment Variables	type	Description
MODULE_NAME	string	Name of the module
MODULE_TYPE	string	Type of the module (Input, Processing, Output)
EGRESS_URLS	string	HTTP ReST endpoints for the next module
INGRESS_HOST	string	Host to which data will be received
INGRESS_PORT	string	Port to which data will be received

Dependencies

bottle
requests
tensorflow-cpu

Input

Input to this module is:

• JSON body single object, example:

{
    "temperature": 12,
    "volume": 1.3,
    "pressure": 0.32
}

• array of JSON body objects, example:

[
    {
        "temperature": 12,
        "volume": 1.3,
        "pressure": 0.32
    },
    {
        "temperature": 13,
        "volume": 2.1,
        "pressure": 0.34
    }
]

Output

Output of this module is

• JSON body single object, example:

{
    "prediction": 14.323
}

• array of JSON body objects, example:

[
    {
        "prediction": 14.323
    },
    {
        "prediction": 13.45
    }
]