neuralarts

Type-Generated Search Spaces for ENAS

The Generator(s)

• Convolutions: the convolution-based search space generator is defined in generate_conv.py and can be used to fix or range three convolution properties: kernel size, padding, dilation. The corresponding k_limit, p_limit, and d_limit should be None, positive integer, or an inclusive positive integer range, and should be one of the following combinations.
  1. One None, one fixed integer (like 3), one integer range (like (3,6)).
  2. One None, two integer ranges (like (3,6)).

Experiments

Brute-Force (T1_*)

Our first experimental group consists of search spaces generated by brute-force using two ranges, with the increasing upper ranges 4/8/12/16, on each pair combination of the three parameters. The third parameter will be None meaning it's "free" and determined as a function of the other two.

• The vv in T1_vvuu defines the variables that are ranged. For example kd means the kernel size and dilation are ranged and the padding is deterministic.
• The uu in T1_vvuu defines the upper range of the ranged variables. For example 08 means the upper (inclusive) bound of the experiment is 8.