Search Space Configuration File Structure¶

This document describes the structure and configuration of the YAML file that defines the search space for supported layers and their parameters. To extend the search space, follow the instructions outlined in the README.md file.

TL;DR¶

✅ The YAML configuration defines the available layers and their parameters.

✅ Parameters using lists follow the [start, stop, step] format.

✅ Rule sets ensure valid model architectures.

✅ Extend the search space by adding layers in layer_definitions.py and updating the YAML.

Extending the Search Space¶

To add a new layer to the search space: 1. Implement the layer in neural_architecture_search/src/layer_definitions.py. 2. If needed, place custom operations in neural_architecture_search/src/search_space_modules/. 3. Add the layer configuration to the gene_pool section of the YAML file. 4. Update the rule_set to define layer connectivity.

Gene Pool¶

The gene_pool section defines available layers and their parameters for feature extraction, global pooling, preprocessing, and dense layers. All list-based parameters follow the [start, stop, step] format.

Feature Extraction - 1D¶

Layer Type	Function Name	Filters/Kernel/Pool Size	Strides	Padding	Activation
`C_1D`	`Conv1D`	Filters: [4, 32, 1], Kernel: [1, 5, 1]	[1, 2, 1]	`same`	`relu`
`DC_1D`	`DepthwiseConv1D`	Kernel: [1, 5, 1]	[1, 2, 1]	`same`	`relu`

Feature Extraction - 2D¶

Layer Type	Function Name	Filters/Kernel/Pool Size	Strides	Padding	Activation
`C_2D`	`Conv2D`	Filters: [4, 32, 1], Kernel: [1, 5, 1]	[1, 2, 1]	`same`	`relu`
`DC_2D`	`DepthwiseConv2D`	Kernel: [1, 5, 1]	[1, 2, 1]	`same`	`relu`

Global Pooling¶

Layer Type	Function Name
`GAP_1D`	`GlobalAveragePooling1D()`
`GMP_1D`	`GlobalMaxPooling1D()`

Preprocessing - 2D¶

Layer Type	Function Name	Parameters
`STFT_2D`	`STFT`	`n_fft`: [64, 1024, 16], `hop_length`: [64, 396, 16]

Dense Layers¶

Layer Type	Function Name	Parameters
`D`	`Dense`	Units: [8, 128, 8], Activation: [`relu`]

Rule Set¶

The rule_set defines valid layer connectivity.

1D Layer Rules¶

Layer	Next Possible Layers
`C_1D`	`AP_1D`, `MP_1D`, `BN_1D`, `DC_1D`, `C_1D`, `STFT_2D`

2D Layer Rules¶

Layer	Next Possible Layers
`STFT_2D`	`MAG_2D`