We human beings don't learn new things from scratch. Instead, we take advantage of what we have learned as much as possible, consciously or not. Transfer learning in AI attempts to do the same thing—it's a technique that takes a normally small piece of a big trained model and reuses it in a new model for a related task, without the need to access the large training data and computing resources to train the original model. Overall, transfer learning is still an open problem in AI, since in many situations, what takes human beings only a few examples of trial-and-errors before learning to grasp something new would take AI a lot more time to train and learn. But in the field of image recognition, transfer learning has proven to be very effective.

Modern deep learning models for image recognition are typically deep neural networks...

In the TensorFlow source that we set up in the previous chapter, there's a Python script, tensorflow/examples/image_retraining/retrain.py, that you can use to retrain the Inception v3 or MobileNet models. Before we run the script to retrain the Inception v3 model for our dog breed recognition, we need to first download the Stanford Dogs Dataset (http://vision.stanford.edu/aditya86/ImageNetDogs), which contains images of 120 dog breeds (you only need to download the Images in the link, not the Annotations).

Untar the downloaded dog images.tar file in ~/Downloads, and you should see a list of folders in ~/Downloads/Images, as shown in the following screenshot. Each folder corresponds to one dog breed and contains about 150 images (you don't need to supply explicit labels for images as the folder names are used to label the images...

The stripped and quantized model generated in the previous section is still over 20 MB in size. This is because the pre-built Inception v3 model used for retraining is a large-scale deep learning model, with over 25 million parameters, and Inception v3 was not created with a mobile-first goal.

In June 2017, Google released MobileNets v1, a total of 16 mobile-first deep learning models for TensorFlow. These models are only a few MB in size, with 0.47 million to 4.24 million parameters, still achieving decent accuracy (just a bit lower than Inception v3). See its README for more information: https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md.

The retrain.py script discussed in the previous section also supports retraining based on MobileNet models. Simply run a command like the following:

python tensorflow/examples...

The iOS simple example we see in Chapter 1, Getting Started with Mobile TensorFlow, uses the Inception v1 model. To make the app use our retrained Inception v3 model and MobileNet model to do better dog breed recognition, we need to make a few changes to the app. Let's first see what it takes to use the retrained quantized_stripped_dogs_retrained.pb in the iOS simple app:

1. Double-click the tf_simple_example.xcworkspace file in tensorflow/examples/ios/simple to open the app in Xcode
2. Drag the quantized_stripped_dogs_retrained.pb model file, the dog_retrained_labels.txt label file, and the lab1.jpg image file we used to test the label_image script, and drop to the project's data folder, making sure both Copy items if needed and Add to targets are checked, as shown in the following screenshot:

To use our retrained Inception v3 model and MobileNet model in Android's TF Classify app is also pretty straightforward. Follow the steps here to test both retrained models:

1. Open the sample TensorFlow Android app, located in tensorflow/examples/android, using Android Studio.
2. Drag and drop two retrained models, quantized_stripped_dogs_retrained .pb and dog_retrained_mobilenet10_224.pb as well as the label file, dog_retrained_labels.txt to the assets folder of the android app.
3. Open the file ClassifierActivity.java, to use the Inception v3 retrained model, and replace the following code:

private static final int INPUT_SIZE = 224; 
private static final int IMAGE_MEAN = 117; 
private static final float IMAGE_STD = 1; 
private static final String INPUT_NAME = "input"; 
private static final String OUTPUT_NAME = &quot...

In the earlier version of TensorFlow, adding TensorFlow to your own app was very tedious, requiring the use of the manual build process of TensorFlow and other manual settings. In TensorFlow 1.4, the process is pretty straightforward, but still, detailed steps are not well documented in the TensorFlow website. One other thing that's missing is the lack of documentation on how to use TensorFlow in your Swift-based iOS app; the sample TensorFlow iOS apps are all in Objective-C, calling TensorFlow's C++ APIs. Let's see how we can do better.

First, follow these steps to add TensorFlow with the image classification feature to your...

It turns out that adding TensorFlow to your own Android app is easier than iOS. Let's jump right to the steps:

1. If you have an existing Android app, skip this. Otherwise, in Android Studio, select File | New | New Project... and accept all the defaults before clicking Finish.
2. Open the build.gradle (Module: app) file, and add compile 'org.tensorflow:tensorflow-android:+' inside and at the end of dependencies {...};.
3. Build the gradle file and you'll see libtensorflow_inference.so, the TensorFlow native library that Java code talks to, inside the subfolders of the location app/build/intermediates/transforms/mergeJniLibs/debug/0/lib of your app directory.
4. If this is a new project, you can create the assets folder by first switching to Packages, then by right-mouse clicking the app and selecting New | Folder | Assets Folder...

In this chapter, we first gave a quick overview on what is transfer learning and why we can and should use it to retrain pretrained deep learning image classification models. Then we presented detailed information on how to retrain an Inception v3-based model and a MobileNet model so we can better understand and recognize our best friends. After that, we first showed how to use the retrained models in the TensorFlow sample iOS and Android apps, then gave step-by-step tutorials on how to add TensorFlow to your own iOS app, both Objective-C and Swift-based, and your own Android app.

Now that we have our best friends, with a few nice and clean tricks, covered, we know there are a lot of other guys out there, good or bad. In the next chapter, we'll learn how to be smarter, how to recognize all the interesting objects in a picture, and locate them, on your smart phone...