Quantize Model

netspresso.np_qai.quantizer.NPQAIQuantizer.quantize_model(input_model_path, output_dir, weights_dtype, activations_dtype, options=QuantizeOptions(), job_name=None, calibration_data=None)

Quantize 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 quantized model.	required
weights_dtype	QuantizeDtype	The data type to use for the weights.	required
activations_dtype	QuantizeDtype	The data type to use for the activations.	required
options	Union[QuantizeOptions, str]	The options to use for the quantization.	QuantizeOptions()
job_name	Optional[str]	The name of the job.	None
calibration_data	Union[Dataset, DatasetEntries, str, None]	The calibration data to use for the quantization.	None

Returns:

Type	Description
Union[NPQAIQuantizerMetadata, List[NPQAIQuantizerMetadata]]	Union[NPQAIQuantizerMetadata, List[NPQAIQuantizerMetadata]]: Returns a quantizer metadata object if successful.

Note

For details, see submit_quantize_job in QAI Hub API.

Example

from netspresso import NPQAI
from netspresso.np_qai import Device
from netspresso.np_qai.options import QuantizePrecision

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

quantizer = np_qai.quantizer()

INPUT_MODEL_PATH = "YOUR_INPUT_MODEL_PATH"
OUTPUT_DIR = "YOUR_OUTPUT_DIR"
JOB_NAME = "YOUR_JOB_NAME"
DEVICE_NAME = "YOUR_DEVICE_NAME"

CALIBRATION_DATA = {"images": inputs_array}

quantized_result = quantizer.quantize_model(
    input_model_path=INPUT_MODEL_PATH,
    output_dir=OUTPUT_DIR,
    calibration_data=CALIBRATION_DATA,
    weights_dtype=QuantizePrecision.INT8,
    activations_dtype=QuantizePrecision.INT8,
    job_name=JOB_NAME,
)
print("Quantization task started")

# Monitor task status
while True:
    status = quantizer.get_quantize_task_status(quantized_result.quantize_info.quantize_task_uuid)
    if status.finished:
        quantized_result = quantizer.update_quantize_task(quantized_result)
        print("Quantization task completed")
        break
    else:
        print("Quantization task is still running")

Create Calibration Datasets

Using good, representative input samples for calibration helps improve performance on target hardware and retains model accuracy

import cv2
from glob import glob

import numpy as np
from netspresso.inferencer.preprocessors.base import Preprocessor

IMG_SIZE = 640
preprocess_list = [
    {
        "name": "resize",
        "size": IMG_SIZE,
        "interpolation": "bilinear",
        "max_size": None,
        "resize_criteria": "long",
    },
    {
        "name": "pad",
        "size": IMG_SIZE,
        "fill": 114,
    }
]

preprocessor = Preprocessor(preprocess_list)

DATASET_PATH = "YOUR_DATASET_PATH"
NUM_DATASET = 100
image_paths = glob(f"{DATASET_PATH}/*.jpg")[:NUM_DATASET]

inputs_array = []

for image_path in image_paths:
    img = cv2.imread(image_path)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = preprocessor(img)
    img = np.transpose(img, (0, 3, 1, 2))
    inputs_array.append(img)

>> np.array(inputs_array).shape
(100, 1, 3, 512, 512)