Configuration Parameters

For each mode, common configuration parameters, HKTaskParams, are required to run the task. These parameters are used to define the task, datasets, model, and other settings. Rather than defining separate configuration files for each mode, a single configuration object is used to simplify configuration files and heavy re-use of parameters between modes.

QuantizationParams

Quantization parameters define the quantization-aware training (QAT) and post-training quantization (PTQ) settings. This is used for modes: train, evaluate, export, and demo.

Argument	Type	Opt/Req	Default	Description
enabled	bool	Optional	False	Enable quantization
qat	bool	Optional	False	Enable quantization aware training (QAT)
format	Literal["int8", "int16", "float16"]	Optional	int8	Quantization mode
io_type	str	Optional	int8	I/O type
conversion	Literal["keras", "tflite"]	Optional	keras	Conversion method
debug	bool	Optional	False	Debug quantization
fallback	bool	Optional	False	Fallback to float32

NamedParams

Named parameters are used to provide custom parameters for a given object or callable where parameter types are not known ahead of time. For example, a dataset, 'CustomDataset', may require custom parameters such as 'path', 'label', 'sampling_rate', etc. When a task loads the dataset using name, the task will then unpack the custom parameters and pass them to the dataset initializer.

Argument	Type	Opt/Req	Default	Description
name	str	Required		Named parameters name
params	dict[str, Any]	Optional	{}	Named parameters

import heartkit as hk

class CustomDataset(hk.HKDataset):

    def __init__(self, a: int = 1, b: int = 2) -> None:
        self.a = a
        self.b = b

hk.DatasetFactory.register("custom", CustomDataset)

params = hk.HKTaskParams(
    datasets=[
        hk.NamedParams(
            name="custom",
            params=dict(a=1, b=2)
        )
    ]
)

HKTaskParams

These parameters are supplied to a Task when running a given mode such as train, evaluate, export, or demo. A single configuration object is used to simplify configuration files and heavy re-use of parameters between modes.

Argument	Type	Opt/Req	Default	Description	Mode
name	str	Required	experiment	Experiment name	All
project	str	Required	heartkit	Project name	All
job_dir	Path	Optional	tempfile.gettempdir	Job output directory	All
datasets	list[NamedParams]	Optional		Datasets	All
force_download	bool	Optional	False	Force download datasets	download
dataset_weights	list[float]|None	Optional	None	Dataset weights	train
sampling_rate	int	Optional	250	Target sampling rate (Hz)	All
frame_size	int	Optional	1250	Frame size in samples	All
samples_per_patient	int|list[int]	Optional	1000	# train samples per patient	train
val_samples_per_patient	int|list[int]	Optional	1000	# validation samples per patient	train
test_samples_per_patient	int|list[int]	Optional	1000	# test samples per patient	evaluate
train_patients	float|None	Optional	None	# or proportion of patients for training	train
val_patients	float|None	Optional	None	# or proportion of patients for validation	train
test_patients	float|None	Optional	None	# or proportion of patients for testing	evaluate
val_file	Path|None	Optional	None	Path to load/store pickled validation file	train
test_file	Path|None	Optional	None	Path to load/store pickled test file	evaluate, export
val_size	int|None	Optional	None	# samples for validation	train
test_size	int	Optional	10000	# samples for testing	evaluate, export
num_classes	int	Optional	1	# of classes	All
class_map	dict[int, int]	Optional		Class/label mapping	All
class_names	list[str]|None	Optional	None	Class names	All
resume	bool	Optional	False	Resume training	train
architecture	NamedParams|None	Optional	None	Custom model architecture	train
model_file	Path|None	Optional	None	Path to load/save model file (.keras)	All
use_logits	bool	Optional	True	Use logits output or softmax	Export
weights_file	Path|None	Optional	None	Path to a checkpoint weights to load/save	train
quantization	QuantizationParams	Optional		Quantization parameters	All
lr_rate	float	Optional	0.001	Learning rate	train
lr_cycles	int	Optional	3	Number of learning rate cycles	train
lr_decay	float	Optional	0.9	Learning rate decay	train
label_smoothing	float	Optional	0	Label smoothing	train
batch_size	int	Optional	32	Batch size	train
buffer_size	int	Optional	100	Buffer cache size	train
epochs	int	Optional	50	Number of epochs	train
steps_per_epoch	int	Optional	10	Number of steps per epoch	train
val_steps_per_epoch	int	Optional	10	Number of validation steps	train
val_metric	Literal["loss", "acc", "f1"]	Optional	loss	Performance metric	train
class_weights	Literal["balanced", "fixed"]	Optional	fixed	Class weights	train
threshold	float|None	Optional	None	Model output threshold	evaluate, export
val_metric_threshold	float|None	Optional	0.98	Validation metric threshold	export
test_metric_threshold	float|None	Optional	0.98	Test metric threshold	export
tflm_var_name	str	Optional	g_model	TFLite Micro C variable name	export
tflm_file	Path|None	Optional	None	Path to copy TFLM header file (e.g. ./model_buffer.h)	export
backend	str	Optional	pc	Backend	demo
demo_size	int|None	Optional	1000	# samples for demo	demo
display_report	bool	Optional	True	Display report	demo
seed	int|None	Optional	None	Random state seed	All
data_parallelism	int	Optional	os.cpu_count	# of data loaders running in parallel	All
verbose	int	Optional	1	Verbosity level	All