Here is a brief overview of the frameworks and principles we use to make OSAI work as efficiently and accurately as possible.


#Frameworks we use

The OSAI service allows the use of predefined architectures that are known to work well, such as ResNet, AlexNet, VGG, Inception, DenseNet, etc. The general scheme of the available models is similar to CNNs (Convolutional Neural Networks), wherein in the first step we have a feature extractor and then a classifier. This directly forces us to apply the transfer learning technique.


#Initializing the model

Neural network models are usually initialized with random weights, which after a series of epochs reach certain values that allow us to correctly classify the input image or detect an object in it. It is also possible to build a baseline model that we will later try to improve during re-training.


#Model re-training

Re-training in OSAI allows you to initialize weights on values that you already know (e.g. from previous training) and that you know will be good for classifying a given dataset. With re-training, we do not need a huge dataset (hundreds of thousands of images or more), because when training a network from scratch, we can be satisfied with just a few thousand. Another advantage, besides the size of the dataset, is fewer epochs for the weights. They will take on good values for classification much faster due to their initialization.



OSAI provides state-of-the-art auto-tuning for better model accuracy.


#Using templates

In OSAI, you can create your own training templates to streamline your work – learn more in the Model Templates section.