Search space documentation#
The search space design is based on fundamental elements originally proposed within an hyper-parameters optimization package called zellij. The search space is made of mix-variables such as integers, floats, categorical variables and graphs. Each of this object is associated with a Variable, which defines what values an object can take.
For example, an integer object will be associated with the variable IntVar, that will take as arguments the lower and upper bounds, defining where the integer is defined.
from dragon.search_space.zellij_variables import IntVar
v = IntVar("An integer variable", 0, 5)
In this example, the variable v defines an integer which can take values from 0 to 5.
These fundamental elements have been leveraged within the DRAGON package to generate new tools for optimizing both the architecture and the hyperparameters of deep neural networks. These tools are very generic and allow the user to use any nn.Module object within the optimized architectures. Some basic operations are already implemented and ready to use to facilitate the use of the package.
The cell-based tutorial shows that DRAGON can be constrained to represent a cell-based search space.