Tensorflow

Tensorflow

Tensorflow Filter allows running Machine Learning inference tasks on the records of data coming from input plugins or stream processor. This filter uses Tensorflow Lite as the inference engine, and requires Tensorflow Lite shared library to be present during build and at runtime.

Tensorflow Lite is a lightweight open-source deep learning framework that is used for mobile and IoT applications. Tensorflow Lite only handles inference (not training), therefore, it loads pre-trained models (.tflite files) that are converted into Tensorflow Lite format (FlatBuffer). You can read more on converting Tensorflow models here

Configuration Parameters

The plugin supports the following configuration parameters:

Key Description Default
input_field Specify the name of the field in the record to apply inference on.
model_file Path to the model file (.tflite) to be loaded by Tensorflow Lite.
include_input_fields Include all input filed in filter’s output True
normalization_value Divide input values to normalization_value

Creating Tensorflow Lite shared library

Clone Tensorflow repository, install bazel package manager, and run the following command in order to create the shared library:

  1. $ bazel build -c opt //tensorflow/lite/c:tensorflowlite_c  # see https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/c

The script creates the shared library bazel-bin/tensorflow/lite/c/libtensorflowlite_c.so. You need to copy the library to a location (such as /usr/lib) that can be used by Fluent Bit.

Building Fluent Bit with Tensorflow filter plugin

Tensorflow filter plugin is disabled by default. You need to build Fluent Bit with Tensorflow plugin enabled. In addition, it requires access to Tensorflow Lite header files to compile. Therefore, you also need to pass the address of the Tensorflow source code on your machine to the build script:

  1. cmake -DFLB_FILTER_TENSORFLOW=On -DTensorflow_DIR=<AddressOfTensorflowSourceCode> ...

Command line

If Tensorflow plugin initializes correctly, it reports successful creation of the interpreter, and prints a summary of model’s input/output types and dimensions.

  1. $ bin/fluent-bit -i mqtt -p 'tag=mqtt.data' -F tensorflow -m '*' -p 'input_field=image' -p 'model_file=/home/user/model.tflite' -p 'include_input_fields=false' -p 'normalization_value=255' -o stdout
  2. [2020/08/04 20:00:00] [ info] Tensorflow Lite interpreter created!
  3. [2020/08/04 20:00:00] [ info] [tensorflow] ===== input #1 =====
  4. [2020/08/04 20:00:00] [ info] [tensorflow] type: FLOAT32  dimensions: {1, 224, 224, 3}
  5. [2020/08/04 20:00:00] [ info] [tensorflow] ===== output #1 ====
  6. [2020/08/04 20:00:00] [ info] [tensorflow] type: FLOAT32  dimensions: {1, 2}

Configuration File

  1. [SERVICE]
  2.     Flush        1
  3.     Daemon       Off
  4.     Log_Level    info
  6. [INPUT]
  7.     Name mqtt
  8.     Tag mqtt.data
  10. [FILTER]
  11.     Name tensorflow
  12.     Match mqtt.data
  13.     input_field image
  14.     model_file /home/m/model.tflite
  15.     include_input_fields false
  16.     normalization_value 255
  18. [OUTPUT]
  19.     Name stdout
  20.     Match *

Limitations

  1. Currently supports single-input models
  2. Uses Tensorflow 2.3 header files