Hello World Example

This example is designed to demonstrate the absolute basics of using TensorFlow Lite for Microcontrollers. It includes the full end-to-end workflow of training a model, converting it for use with TensorFlow Lite for Microcontrollers for running inference on a microcontroller.

Table of contents

Run the evaluate.py script on a development machine
Run the tests on a development machine
Train your own model

Run the evaluate.py script on a development machine

The evaluate.py script runs the hello_world.tflite model with x_values in the range of [0, 2*PI]. The script plots a diagram of the predicted value of sinwave using TFLM interpreter and compare that prediction with the actual value generated by the numpy lib.

bazel build tensorflow/lite/micro/examples/hello_world:evaluate
bazel run tensorflow/lite/micro/examples/hello_world:evaluate
bazel run tensorflow/lite/micro/examples/hello_world:evaluate -- --use_tflite

Run the evaluate_test.py script on a development machine

These tests verify the input/output as well as the prediction of the hello_world.tflite model. There is a test to also verify the correctness of the model by running both TFLM and TFlite interpreter and then comparing the prediction from both interpreters.

bazel build tensorflow/lite/micro/examples/hello_world:evaluate_test
bazel run tensorflow/lite/micro/examples/hello_world:evaluate_test

Run the tests on a development machine

Run the cc test using bazel

bazel run tensorflow/lite/micro/examples/hello_world:hello_world_test

And to run it using make

make -f tensorflow/lite/micro/tools/make/Makefile test_hello_world_test

The source for the test is hello_world_test.cc. It's a fairly small amount of code that creates an interpreter, gets a handle to a model that's been compiled into the program, and then invokes the interpreter with the model and sample inputs.

Train your own model

So far you have used an existing trained model to run inference on microcontrollers. If you wish to train your own model, here are the scripts that can help you to achieve that.

bazel build tensorflow/lite/micro/examples/hello_world:train

And to run it

bazel-bin/tensorflow/lite/micro/examples/hello_world/train --save_tf_model 
--save_dir=/tmp/model_created/

The above script will create a TF model and TFlite model inside the /tmp/model_created directory.

Now the above model is a float model. Means it can take floating point input and can produce floating point output.

If we want a fully quantized model we can use the ptq.py script inside the quantization directory. The ptq.py script can take a floating point TF model and can produce a quantized model.

Build the ptq.py script like

bazel build tensorflow/lite/micro/examples/hello_world/quantization:ptq

Then we can run the ptq script to convert the float model to quant model as follows. Note that we are using the directory (/tmp/model_created) of the TF model as the source_model_dir here. The quant model (named hello_world_int8.tflite) will be created inside the target_dir. The ptq.py script will convert the TF model found inside the /tmp/model_created folder and convert it to a int8 TFlite model.

bazel-bin/tensorflow/lite/micro/examples/hello_world/quantization/ptq  
--source_model_dir=/tmp/model_created --target_dir=/tmp/quant_model/