Skip to main content

Getting started

So you have decided to give Alvíss AI a go? Great! This guide will introduce you to the basic concepts on our platform and how to create a model.

Basic concepts

Everything you do in Alvíss AI is considered a job. A job can be initiating a model, training, predicting, or other regular actions in building and using machine learning models. Some types of jobs create a model, which you can then use in other jobs.

A capability is a module that solve a type of problems e.g. object dectection, OCR, time series, etc. Each capability defines what functions are available. Typically, all capabilities include create model, train and predictfunctions.

How you will be using it

The simplest workflow is to create a pre-trained model and directly use it to predict on your data. We offer a range of models trained with well-known datasets e.g. VOC, CIFAR, and more. With our pre-trained models, you get a decent prediction within a few clicks.

As an example, here is the quickest way to use object detection:

  1. Create a model with pre-trained weights with pretrained model function. Try create one now.
  2. Use the model (job) you just created and send in an image to predict function. You then get a prediction.

As simple as that.

Simple 2-step workflow to get object detection prediction

If pre-trained models does not suffice, either because you have domain-specific case or merely want to push the performance even further, you can train your own model. You have an option to start from one of the pre-trained models, which can potentially reach your target accuracy with less training time. You can also create a new model and train it from scratch. With this approach, you may choose a network architecture or tune some parameters. We provide defaults value that present best average based on our researches and experiments, but you have full control of tweaking the configuration as you want.

In this example, a pre-trained model is created and then trained with a new dataset. The training job produces a new model which is used in evaluate and predict. When a capability supports incremental training, you can iterate back to training until you get satisfactory results.

Typical workflow in developing a new machine learning model

Note that each capability defines its own functions. Some of them may create network within train function, for example, in which case the first step in the diagram above is not explicit.