Featrix Support: Getting Started
Getting Started with Featrix: Creating a Neural Function
The neural function is the basis of everything you do in Featrix. Let's take a look.
Creating a Project and
Connecting Data
Creating a neural function involves two things: First, you will create a project and connect some data. You can work with data that you upload, download from a URL, a public data source, and more.
For this example, we will work with data that Featrix provides. This is a credit data set that is part of the OpenML data available in sklearn.
We'll start by clicking 'new project' and then clicking 'create new API project'
Next, we'll click 'connect new data' in our project window, and then click 'enter public URL to load'. When we click this, a box pops up and we will enter the following URL:
https://bits.featrix.com/demo-data/github.com-anujtiwari21/train.csv
Once we load the data, Featrix will take a look at it and evaluate what we have. After a few seconds, the project screen will contain information about the columns in the data, the types detected, and provide controls for us to ignore (not include) data in the trained functions.
Creating a Neural Function
We're now ready to create the neural function. We hit the "Create New Neural Function" button and we get a screen asking us to define the target field and pick a training budget. The target field is the field we want to predict with our model. In this case, it's the field as to whether or not we want to make a loan to a customer.
The training budget sets an upper limit on how much you'll spend on training the function. This lets you keep costs under control. You can always apply more budget to the function later without losing your initial investment. For this model, 3 to 5 credits is enough to get excellent performance.
If Featrix detects the model will not improve by more work, then you may come in under budget, but we'll never charge you more than your budget for building the function.
Once you're ready to go, hit the "Let's Build It" button. Training can take a few minutes.
Training the Neural Function
The training happens in two steps: First we train an underlying foundation that you can build multiple functions on. Then, we train your specific function. If you already have trained a foundation, we'll use that and just build the function on top of the foundation.
Once your function is ready, it will show its status as "Ready to run".
Ad hoc queries in the User Interface
You can play with the neural function right from the console by hitting the "sandbox" button.
The sandbox lets you type in partial or full queries to your function with no transformations or encoding required on your part. You can put in the same symbols that you trained the model with, including strings symbols.
Once you have a valid JSON object, Run Prediction will execute your model. When the model returns an answer, you'll get another JSON object in a separate field called "Result".
Notes:
- Manually typing a JSON object isn't fun. We'll make this friendlier in an upcoming release.
- You can enter either a JSON dictionary for one query, or a list of dictionaries to run several queries in bulk.
- You do not have to specify an entire record: you can input as little as one field. Of course, the more inputs you provide, the better the answer will fit the original data.
- This execution can be a little slower than in production because of the ad hoc nature of the query. If you use the function over and over in production, our backend will cache your models and performance will be higher.
Queries with the API
To integrate the neural function into your application, you can use the "show code" button to get a template for Python calls to our API.
Our API does everything you see in the UI and more.
If you want JavaScript, TypeScript, Swift, C#, Ruby, etc. integrations, please get in touch. We are working on more language support natively. The calls are HTTPS REST under the hood.
Working with Featrix in a Jupyter Notebook is easy and fully supported.
Building another function
Once you have a first function on a specific target field, you can build a second function for a different field and use the existing foundation base you already have build. All you need to do to add a second function is click the "Create New Neural Function" button and add a second function.
Training More
You can build additional versions of your functions by doing more training. You can train all of your functions more or just a subset.
Working with Multiple Data Sets
Connecting data together is often a lot of work and we've simplified it a lot with Featrix. We will have a support page about this coming shortly; please write to hello@featrix.ai for help in the interim.
More About Billing
Billing works by credits. Credits are used whenever you train, store, or run a function. You can buy credit packs and add them to your account; we are waiting for our payments provider to enable auto-refilling of credits, which should happen in early Q3.
Q&A
- Do I have to clean my data?
- What about multiple data sets?
- How many credits will it take to get a good result on my data?
- What data types does Featrix support?
- Can I upload my data via API?
- Can I stream events into Featrix?
- My data schema is going to change, how do I deal with that?
- I need to update my model with new data, how do I do that?
Additional Help
These additional articles are here to help.
-
API Keys
Learn about our API keys, security, and teams.
-
API Documentation
Everything in Featrix is accessible via our documented API.
-
Multiple Data Sets
Learn how to connect multiple data sets together, even when there is no join key.
-
Neural Function Performance
Learn about accuracy, precision, recall, and other metrics with Featrix.
-
Updating Data
Learn how to add new versions of data and train new models.
-
Billing and Credits
Learn how billing and credits work.
How can we help?
Reach out via hello@featrix.ai or schedule time to meet with a developer on our team.