Usage

The top-level entry point of the framework is the HiveNAS class.

Examples

The simplest way to run HiveNAS is to use the default parameters (which are empirically optimized for simple image-classification tasks) as follows:

python HiveNAS.py --verbose

To specify a custom configuration file (see Configuration), use the -c | --config-file argument:

python HiveNAS.py --verbose -c="path/to/config.yaml"

To import the top-level module into an existing project:

from HiveNAS import HiveNAS

HiveNAS.find_topology()

All primitive, non-iterable operational parameters (i.e int, float, bool, str) are defined as CLI arguments. Find more customizable use cases below.

Advanced Usage

To run HiveNAS over the MNIST benchmark, specify a unique configuration version name and the dataset:

python HiveNAS.py --verbose -cv="mnist-trial" -oo="NAS" -ds="MNIST"

The same setting over the FashionMNIST benchmark can be ran as follows:

python HiveNAS.py --verbose -cv="fashion-mnist-trial" -oo="NAS" -ds="FASHION_MNIST"

To experiment with Artificial Bee Colony and Numerical Benchmarks (Rosenbrock optimization, for instance), the -oo flag (or --optimization-objective) can be specified.

python HiveNAS.py --verbose -cv="rosenbrock-trial" -oo="Rosenbrock"

CLI Arguments

To override any of the default parameters, refer to the table below (or use the -h / --help argument):

Configuration Parameter	Argument Name	Argument Flag
CONFIG_VERSION	--config-version	-cv
SEED_VALUE	--seed-value	-sv
OPTIMIZATION_OBJECTIVE	--optimization-objective	-oo
ABANDONMENT_LIMIT	--abandonment-limit	-al
COLONY_SIZE	--colony-size	-cs
EMPLOYEE_ONLOOKER_RATIO	--employee-onlooker-ratio	-eor
ITERATIONS_COUNT	--iterations-count	-ic
RESULTS_SAVE_FREQUENCY	--results-save-frequency	-rsf
RESULTS_BASE_PATH	--results-base-path	-rbp
HISTORY_FILES_SUBPATH	--history-files-subpath	-hfs
ENABLE_WEIGHT_SAVING	--enable-weight-saving	-ews
WEIGHT_FILES_SUBPATH	--weight-files-subpath	-wfs
RESUME_FROM_RESULTS_FILE	--resume-from-results-file	-rfrf
DEPTH	--depth	-d
STOCHASTIC_SC_RATE	--stochastic-sc-rate	-ssr
DATASET	--dataset	-ds
EPOCHS	--epochs	-e
MOMENTUM_EPOCHS	--momentum-epochs	-me
FULL_TRAIN_EPOCHS	--full-train-epochs	-fte
TERMINATION_THRESHOLD_FACTOR	--termination-threshold-factor	-ttf
TERMINATION_DIMINISHING_FACTOR	--termination-diminishing-factor	-tdf
INITIAL_LR	--initial-lr	-il
FINAL_LR	--final-lr	-fl
BATCH_SIZE	--batch-size	-bs
AFFINE_TRANSFORMATIONS_ENABLED	--affine-transformations-enabled	-ate
CUTOUT_PROB	--cutout-prob	-cp
SATURATION_AUG_PROB	--saturation-aug-prob	-sap
CONTRAST_AUG_PROB	--contrast-aug-prob	-cap

Alternative Datasets

In addition to the predefined datasets, virtually any labeled type of data can be used by defining a custom loader in NASEval's initializer.

Note

Ensure that the input data (i.e (X_train, X_test)) have a 4-dimensional shape, matching Keras’ CIFAR10 . If the data has less channels, consider adding placeholders as demonstrated in NASEval with MNIST/FashionMNIST (X_train.reshape(-1,28,28,1))

HiveNAS currently supports CNNs only, with plans to expand to RNNs in the future.