export_and_quantize.sh

#!/usr/bin/env bash
# SPDX-License-Identifier: MIT
# Copyright (C) Intel Corporation
#
# Export a YOLO26 detector to OpenVINO IR for the object-classification use case.
# The accompanying samples group detected objects into higher-level city
# categories (People, Vehicles) for traffic situational awareness.
# Usage: ./export_and_quantize.sh [MODEL_VARIANT] [PRECISION]
# Example: ./export_and_quantize.sh yolo26n FP16
#
# Supported precisions:
#   FP32  -- Full-precision floating-point weights
#   FP16  -- Half-precision floating-point weights (default)
#   INT8  -- Quantized 8-bit integer weights (requires NNCF)
#
# Precision / device compatibility:
#   | Precision | CPU | GPU | NPU |
#   |-----------|-----|-----|-----|
#   | FP32      | Yes | Yes | No  |
#   | FP16      | Yes | Yes | Yes |
#   | INT8      | Yes | Yes | Yes |

set -euo pipefail

MODEL_NAME="${1:-yolo26n}"
PRECISION="${2:-FP16}"
PRECISION="$(echo "${PRECISION}" | tr '[:lower:]' '[:upper:]')"

if [[ "${PRECISION}" != "FP32" && "${PRECISION}" != "FP16" && "${PRECISION}" != "INT8" ]]; then
    echo "ERROR: unsupported precision '${PRECISION}'. Choose FP32, FP16, or INT8." >&2
    exit 1
fi

echo "--- Installing dependencies ---"
if [[ "${PRECISION}" == "INT8" ]]; then
    pip install -qU openvino nncf ultralytics
else
    pip install -qU openvino ultralytics
fi

# Ask for approval before downloading models and sample files
echo ""
echo "This script will download:"
echo "  - Model weights and a sample traffic video"
echo ""
read -p "Continue with downloads? (yes/no): " APPROVAL
if [[ "${APPROVAL}" != "yes" ]]; then
    echo "Download cancelled by user."
    exit 0
fi

# Ping the HuggingFace repo to register a tracked download of config.json.
# This is best-effort: a failure (offline, or before the repo is published)
# must not stop the export.
echo "--- Registering HuggingFace download (tracking ping) ---"
HF_REPO_ID="Intel/object-classification"
HF_CONFIG_URL="https://huggingface.co/${HF_REPO_ID}/resolve/main/config.json"
if curl -fsSL -o /dev/null "${HF_CONFIG_URL}"; then
    echo "Registered HuggingFace download for ${HF_REPO_ID}"
else
    echo "WARNING: HuggingFace tracking ping failed (offline?); continuing." >&2
fi

echo ""
echo "--- Downloading sample test video (urban roundabout) ---"
if [[ ! -f test_video.mp4 ]]; then
    wget -q -O test_video.mp4 \
        "https://www.pexels.com/download/video/30119018/?fps=59.94&h=720&w=1280"
    echo "Downloaded: test_video.mp4"
else
    echo "Already present: test_video.mp4"
fi

if [[ "${PRECISION}" == "FP32" ]]; then
    HALF_FLAG="False"
    EXPORT_LABEL="FP32"
else
    HALF_FLAG="True"
    EXPORT_LABEL="FP16"
fi

echo "--- Exporting ${MODEL_NAME} to OpenVINO IR (${EXPORT_LABEL}) ---"
python3 -c "
from ultralytics import YOLO

model = YOLO('${MODEL_NAME}.pt')
model.export(format='openvino', half=${HALF_FLAG}, dynamic=False, imgsz=640)
print('Export complete: ${MODEL_NAME}_openvino_model/')
"

if [[ "${PRECISION}" == "INT8" ]]; then
    echo "--- Quantizing to INT8 with NNCF ---"
    python3 -c "
import nncf
import openvino as ov
import numpy as np
import cv2

core = ov.Core()
model = core.read_model('${MODEL_NAME}_openvino_model/${MODEL_NAME}.xml')

# Calibrate on a representative frame from the sample traffic video.
cap = cv2.VideoCapture('test_video.mp4')
ok, frame = cap.read()
cap.release()
if not ok:
    raise SystemExit('Could not read a calibration frame from test_video.mp4')
img = cv2.resize(frame, (640, 640))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
img = img.transpose(2, 0, 1)[np.newaxis, ...]  # NCHW

def transform_fn(data_item):
    return img

calibration_dataset = nncf.Dataset(list(range(300)), transform_fn)

quantized = nncf.quantize(
    model,
    calibration_dataset,
    preset=nncf.QuantizationPreset.MIXED,
    subset_size=300,
)

ov.save_model(quantized, '${MODEL_NAME}_objcls_int8.xml')
print('Quantization complete: ${MODEL_NAME}_objcls_int8.xml')
"
fi
echo "--- Done ---"