Convert Model

netspresso.np_qai.converter.NPQAIConverter.convert_model(input_model_path, output_dir, target_device_name, input_shapes=None, options=CompileOptions(), job_name=None, single_compile=True, calibration_data=None, retry=True)

Convert a model in the QAI hub.

Parameters:

Name	Type	Description	Default
input_model_path	Union[str, Path]	The path to the input model.	required
output_dir	str	The directory to save the converted model.	required
target_device_name	Union[Device, List[Device]]	The device to compile the model for.	required
input_shapes	Optional[InputSpecs]	The input shapes of the model.	None
options	Union[CompileOptions, str]	The options to use for the conversion.	CompileOptions()
job_name	Optional[str]	The name of the job.	None
single_compile	bool	Whether to compile the model in a single step.	True
calibration_data	Union[Dataset, DatasetEntries, str, None]	The calibration data to use for the conversion.	None
retry	bool	Whether to retry the conversion if it fails.	True

Returns:

Type	Description
Union[ConverterMetadata, List[ConverterMetadata]]	Union[ConverterMetadata, List[ConverterMetadata]]: Returns a converter metadata object if successful.

Note

For details, see submit_compile_job in QAI Hub API.

Example

from netspresso import NPQAI
from netspresso.np_qai import Device
from netspresso.np_qai.options import CompileOptions, Runtime, ComputeUnit, QuantizeFullType

QAI_HUB_API_TOKEN = "YOUR_QAI_HUB_API_TOKEN"
np_qai = NPQAI(api_token=QAI_HUB_API_TOKEN)

converter = np_qai.converter()

convert_options = CompileOptions(
    target_runtime=Runtime.TFLITE,
    compute_unit=[ComputeUnit.NPU],
    quantize_full_type=QuantizeFullType.INT8,
    quantize_io=True,
    quantize_io_type=QuantizeFullType.INT8,
)

IMG_SIZE = 640
INPUT_MODEL_PATH = "YOUR_INPUT_MODEL_PATH"
OUTPUT_DIR = "YOUR_OUTPUT_DIR"
JOB_NAME = "YOUR_JOB_NAME"
DEVICE_NAME = "QCS6490 (Proxy)"
converted_result = converter.convert_model(
    input_model_path=INPUT_MODEL_PATH,
    output_dir=OUTPUT_DIR,
    target_device_name=Device(DEVICE_NAME),
    options=convert_options,
    input_shapes=dict(image=(1, 3, IMG_SIZE, IMG_SIZE)),
    job_name=JOB_NAME,
)

print("Conversion task started")

# Monitor task status
while True:
    status = converter.get_convert_task_status(converted_result.convert_task_info.convert_task_uuid)
    if status.finished:
        converted_result = converter.update_convert_task(converted_result)
        print("Conversion task completed")
        break
    else:
        print("Conversion task is still running")