Dataset Configuration File Structure¶

This document provides instructions on how to structure and configure the YAML file for each dataset. The configuration file contains general parameters, dataset-specific settings, neural network training hyperparameters, evolutionary search parameters, and fitness evaluation criteria.

TL;DR¶

✅ The YAML configuration defines dataset parameters and hyperparameters.

✅ The tables provide an overview of each parameter and its purpose.

✅ Example YAML snippets illustrate proper formatting.

✅ Decay lists dynamically adjust hyperparameters across generations to optimize the search process.

General Parameters¶

These parameters define general settings that apply to the optimization and hardware constraints.

Parameter	Value Type	Description
`min_free_space_gpu`	Integer	Minimum GPU memory (in bytes) required for execution.

Example:

min_free_space_gpu:
  value: 6_000_000

Dataset/DNN Training Hyperparameters¶

These parameters define dataset-related and deep neural network training configurations.

Parameter	Value Type	Description
`results_path`	String	Directory where training results are stored.
`dataset_name`	String	Name of the dataset used for training.
`sample_rate`	Integer	Sample rate of the dataset (e.g., in Hz).
`input_shape`	List	Input shape of the data.
`num_classes`	Integer	Number of output classes.
`top_activation`	String	Activation function for the output layer.
`num_epochs`	Integer	Number of training epochs.
`batch_size`	Integer	Batch size for training.
`optimizer`	String	Optimizer used for training.
`loss`	String	Loss function used in training.
`metrics`	List	Evaluation metrics used during training.

Example:

results_path:
  value: "Results/"
dataset_name:
  value: "speech_commands"
sample_rate:
  value: 6000
input_shape:
  value: [6000, 1]
num_classes:
  value: 12
top_activation:
  value: "softmax"
num_epochs:
  value: 3
batch_size:
  value: 128
optimizer:
  value: "adam"
loss:
  value: "categorical_crossentropy"
metrics:
  value: ["accuracy"]

EdgeVolution Hyperparameters¶

These parameters define the settings for evolutionary neural architecture search (EdgeVolution). The decay lists specify how hyperparameters change across generations to balance exploration and exploitation.

Parameter	Value Type	Description
`num_generations`	Integer	Total number of generations for the search process.
`population_size_decay`	List	Defines how the population size reduces over generations.
`num_best_models_crossover_decay`	List	Determines how many top models participate in crossover as generations progress.
`mutation_rate_decay`	List	Specifies the mutation rate adjustments at different generations.
`max_num_feature_layers`	Integer	Maximum number of feature extraction layers allowed.
`max_num_classification_layers`	Integer	Maximum number of classification layers allowed.

Explanation of Decay Lists: - Each list consists of tuples in the format [(generation, value), ...]. - The value takes effect from the specified generation onwards. - This allows gradual tuning of parameters as training progresses.

Example:

num_generations:
  value: 30
population_size_decay:
  value: [[1, 500], [2, 250], [10, 100]]
num_best_models_crossover_decay:
  value: [[1, 100], [2, 50], [10, 20]]
mutation_rate_decay:
  value: [[1, 30], [2, 25], [5, 20], [10, 15]]
max_num_feature_layers:
  value: 8
max_num_classification_layers:
  value: 4

Fitness Evaluation Criteria¶

These parameters define the fitness function used in model selection.

Parameter	Value Type	Description
`min_rom_usage`	Integer	Minimum required ROM usage in bytes.
`min_energy_information`	Float	Minimum energy efficiency required.
`acc_weight`	Float	Weight for accuracy in fitness evaluation.
`rom_usage_weight`	Float	Weight for ROM usage in fitness evaluation.
`energy_information_weight`	Float	Weight for energy efficiency in fitness evaluation.

Example:

min_rom_usage:
  value: 150_000
min_energy_information:
  value: 0.5
acc_weight:
  value: 0.7
rom_usage_weight:
  value: 0.1
energy_information_weight:
  value: 0.2