In this lesson, you will build a classification model using some of the techniques you have learned in previous lessons and with the delicious cuisine dataset used throughout this series. In addition, you will build a small web app to use a saved model, leveraging Onnx's web runtime.
[](https://youtu.be/giIXNoiqO_U "Recommendation Systems Introduction")
Building applied ML systems is an important part of leveraging these technologies for your business systems. You can use models within your web applications (and thus use them in an offline context if needed) by using Onnx. In a [previous lesson](../../3-Web-App/1-Web-App/README.md), you built a Regression model about UFO sightings, "pickled" it, and used it in a Flask app. While this architecture is very useful to know it is a full-stack Python app, and your requirements may include the use of a JavaScript application. In this lesson, you can build a basic JavaScript-based system for inference. First, however, you need to train a model and convert it for use with Onnx.
Now, convert your model to Onnx. Make sure to do the conversion with the proper Tensor number. This dataset has 380 ingredients listed, so you need to notate that number in `FloatTensorType`:
```python
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
> Note, you can pass in [options](https://onnx.ai/sklearn-onnx/parameterized.html) in your conversion script. In this case, we passed in 'nocl' to be True and 'zipmap' to be False. Since this is a classification model, you have the option to remove ZipMap which produces a list of dictionaries (not necessary). `nocl` refers to class information being included in the model. Reduce your model's size by setting `nocl` to 'True'.
Onnx models are not very visible in Visual Studio code, but there's a very good free software that many researchers use to visualize the model to ensure that it is properly built. Download [Netron](https://github.com/lutzroeder/Netron) and open your model.onnx file. You can see your simple model visualized, with its 380 inputs and classifier listed:
Now you are ready to use this neat model in a web app. Let's build an app that will come in handy when you look in your refrigerator and try to figure out which combination of your leftover ingredients you can use to cook a given cuisine, as determined by your model.
## Build a recommender web application
You can use your model directly in a web app. This architecture also allows you to run it locally and even offline if needed. Start by creating an `index.html` file in the same folder where you stored your `model.onnx` file.
<buttononClick="startInference()">What kind of cuisine can you make?</button>
</div>
```
Notice that each checkbox is given a value. This reflects the index where the ingredient is found according to the dataset. Apple, for example, in this alphabetic list, occupies the fifth column, so its value is '4' since we start counting at 0. You can consult the [ingredients spreadsheet](../data/ingredient_indexes.csv) to discover a given ingredient's index.
Continuing your work in the index.html file, add a script block where the model is called after the final closing `</div>`. First, import the [Onnx Runtime](https://www.onnxruntime.ai/):
const input = new ort.Tensor(new Float32Array(ingredients), [1, 380]);
const feeds = { float_input: input };
// feed inputs and run
const results = await session.run(feeds);
// read from results
alert('You can enjoy ' + results.label.data[0] + ' cuisine today!')
} catch (e) {
console.log(`failed to inference ONNX model: ${e}.`);
}
}
else alert("Please check an ingredient")
}
init();
</script>
```
In this code, there are several things happening:
1. You created an array of 380 possible values (1 or 0) to be set and sent to the model for inference, depending on whether an ingredient checkbox is checked.
2. You created an array of checkboxes and a way to determine whether they were checked in an `init` function that is called when the application starts. When a checkbox is checked, the `ingredients` array is altered to reflect the chosen ingredient.
3. You created a `testCheckboxes` function that checks whether any checkbox was checked.
4. You use that function when the button is pressed and, if any checkbox is checked, you start inference.
5. The inference routine includes:
1. Setting up an asyncronous load of the model
2. Creating a Tensor structure to send to the model
3. Creating 'feeds' that reflects the `float_input` input that you created when training your model (you can use Netron to verify that name)
4. Sending these 'feeds' to the model and waiting for a response
## Test your application
Open a terminal session in Visual Studio Code in the folder where your index.html file resides. Ensure that you have `[http-server](https://www.npmjs.com/package/http-server)` installed globally, and type `http-server` at the prompt. A localhost should open and you can view your web app. Check what cuisine is recommended based on various ingredients:
Congratulations, you have created a simple web app recommendation with a few fields. Take some time to build out this system!
## 🚀Challenge
Your web app is very minimal, so continue to build it out using ingredients and their indexes from the [ingredient_indexes](../data/ingredient_indexes.csv) data. What flavor combinations work to create a given national dish?
While this lesson just touched on the utility of creating a recommendation system for food ingredients, this area of ML applications is very rich in examples. Read some more about how these systems are built:
