Over a million developers have joined DZone.

How to Deploy TensorFlow Models on Edge Devices

DZone's Guide to

How to Deploy TensorFlow Models on Edge Devices

Want to learn how to deploy TensorFlow Models on your edge devices? Check out this post to learn how.

· IoT Zone ·
Free Resource

While it has been possible to deploy TensorFlow models to mobile and embedded devices via TensorFlow Mobile for some time, Google released an experimental version of TensorFlow Lite as an evolution of TensorFlow Mobile at the end of last year. This new functionality helps to build exciting AI scenarios on edge devices. Not only that, the performance of the model is amazing!

This article, along with my code on GitHub, describes how to train models via the IBM Watson Studio in the cloud or locally. After training, the models are optimized so that they can be deployed to various devices, for example, iOS and Android-based systems.

For an example, I picked a Visual Recognition scenario that is similar to my earlier blog entry where I described how to use TensorFlow.js in browsers. In order to train the model, I’ve taken pictures of seven items: a plug, a soccer ball, mouse, hat, truck, a banana, and headphones.

Check out this video for a quick demo of how these items can be recognized in iOS and Android apps. This is an iPhone screenshot showing a banana being recognized.


Google provides a nice and easy tutorial on how to train, optimize, and deploy a Visual Recognition model to Android devices. I’ve tried to deploy the model on an iPhone, but I have ran into some issues. Check out my project on GitHub to see how I’ve fixed those issues.

  • Since TensorFlow Lite is only experimental, interfaces have changed. To follow the Google tutorial, you need to use the exact TensorFlow version 1.7 and not the later ones.
  • For me, it wasn’t easy to install and run the optimization tool. I ended up using a Docker image which comes with TensorFlow and the pre-compiled tools.
  • The tutorial describes how to generate an optimized model without quantization. The iOS sample app, however, expects a quantized model.

In order to efficiently run models on edge devices, they need to be optimized in terms of model size, memory usage, battery usage, etc. To achieve this, one method is to remove dropouts from the graph. Dropouts are used during training to prevent overfitting models. At runtime, when running predictions, they are not needed.

Another method to optimize models is quantization. Weights in graphs are often defined via floats. When using integers, the sizes of the models are reduced significantly while the accuracy is only affected minimally or not at all.

In order to get the iOS camera sample working with the model that is generated in the tutorial, I had to do some changes. The integers had to be changed to floats when passing in an image to the graph:

const float input_mean = 0.0f;
const float input_std = 255.0f;

float* out = interpreter->typed_input_tensor<float>(0);
for (int y = 0; y < wanted_input_height; ++y) {
  float* out_row = out + (y * wanted_input_width * wanted_input_channels);
  for (int x = 0; x < wanted_input_width; ++x) {
    const int in_x = (y * image_width) / wanted_input_width;
    const int in_y = (x * image_height) / wanted_input_height;
    uint8_t* in_pixel = in + (in_y * image_width * image_channels) + (in_x * image_channels);
    float* out_pixel = out_row + (x * wanted_input_channels);
    for (int c = 0; c < wanted_input_channels; ++c) {
      out_pixel1 = (in_pixel1 - input_mean) / input_std;

After this prediction, it can be invoked in the Objective-C++ code to look like this:

if (interpreter->Invoke() != kTfLiteOk) {
  LOG(FATAL) << "Failed to invoke!";
float* output = interpreter->typed_output_tensor<float>(0); 
GetTopN(output, output_size, kNumResults, kThreshold, &top_results);
NSMutableDictionary* newValues = [NSMutableDictionary dictionary];
for (const auto& result : top_results) {
  const float confidence = result.first;
  const int index = result.second;
  NSString* labelObject = [NSString stringWithUTF8String:labels[index].c_str()];
  NSNumber* valueObject = [NSNumber numberWithFloat:confidence];
  [newValues setObject:valueObject forKey:labelObject];

If you want to run this example yourself, get the code from GitHub and a free IBM Cloud account.

iot ,tensorflow ,edge devices ,tutorial ,ibm watson ,tensorflow lite ,visual recognition

Published at DZone with permission of

Opinions expressed by DZone contributors are their own.

{{ parent.title || parent.header.title}}

{{ parent.tldr }}

{{ parent.urlSource.name }}