Going further with Image Classification

Key Takeaways

[Music] so you've built an image classifier app that recognizes the contents of an image using the generic model it works really well but you probably want to make it more specific so instead of recognizing this image as a plant with petals you want to go further and have it recognize it as a daisy or this as a rose and these as tulips although maybe similar scenarios maybe to identify different types of bird or dessert or whatever you're going to need your own model to do this so if you feed it in a daisy you'll get back a set of predictions for that flower and recall that the output of the neural network will be the number of values based on the classes that the model is trained for so for example here i have three outputs for the probability of a daisy a rose or a tulip so how would you build a model like this well i'm glad you asked the first thing you'll need to do is collect samples of each of the classes that you want to detect so for example in the codelab later i'll provide a data set with roses daisies tulips sunflowers and dandelions each of these classes are in a folder so all of the roses will be in a folder called roses and all of the daisies in a folder called daisies and so on these folders are then zipped up into what we'll call our data set of flower images then you'll use google colab which is a browser-based development environment that lets you code in python but you don't need to worry about it installing python or any of its dependencies you'll also get powerful gpu-based hardware on the back end to speed up your neural network training this will let you code with a tool called tensorflow lite model maker and its purpose is to simplify the process of coding and training a neural network while encapsulating many of the steps that you'd normally have to write yourself and if you're not a pythonista yet it's okay because this does this in a simple high-level api which greatly reduces the complexity into a piece of code like this and here you can create an image classifier data loader to load in the images from their directories your data scientist friends probably told you not to use all of your data to train the network and you should hold a little bit back for testing later well you can do that by just splitting the data like we do here with 90 of it being used for training and the other 10 being held back and then you can just create an image classifier using the data and then output a model for tensorflow lite that can be used on android like this once you have your model you can then integrate it into your app using ml kit's custom model apis and all of this will be covered in the next two codelabs but before you get to them let's look at your expected results here's the original app that you built using the base model in ml kit and when you try to label the image i see things like pedal flower plant and sky but when i use my own model and update the app with that and then label the image i can see that this image is a daisy with 95.9 probability similarly in ios the original app with the base model if i classify the flower i'll see paddle flower sky and plants but when i update the app with the custom model for flowers and i try that i see that it also detects that this image is a daisy so in the next two code labs you'll first train the model and then you'll see how to integrate it into your app using mlkit's custom models so let's have some fun coding you