infer_onnx.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import os
import sys
import cv2
import pathlib
import argparse
import numpy as np
import onnxruntime as ort

CURRENT_DIR = pathlib.Path(__file__).parent
sys.path.append(str(CURRENT_DIR))

from coco import COCO_CLASSES
from demo_utils import mkdir, multiclass_nms, demo_postprocess, vis


def make_parser():
    parser = argparse.ArgumentParser("onnxruntime inference sample")
    parser.add_argument(
        "-m",
        "--model",
        type=str,
        default="yolox-s-int8.onnx",
        help="Input your onnx model.",
    )
    parser.add_argument(
        "-i",
        "--image_path",
        type=str,
        default='test_image.png',
        help="Path to your input image.",
    )
    parser.add_argument(
        "-o",
        "--output_dir",
        type=str,
        default='demo_output',
        help="Path to your output directory.",
    )
    parser.add_argument(
        "-s",
        "--score_thr",
        type=float,
        default=0.3,
        help="Score threshold to filter the result.",
    )
    parser.add_argument(
        "--input_shape",
        type=str,
        default="640,640",
        help="Specify an input shape for inference.",
    )
    parser.add_argument(
        "--ipu",
        action="store_true",
        help="Use IPU for inference.",
    )
    parser.add_argument(
        "--provider_config",
        type=str,
        default="vaip_config.json",
        help="Path of the config file for setting provider_options.",
    )
    return parser


def preprocess(img, input_shape, swap=(2, 0, 1)):
    """
    Preprocessing part of YOLOX for scaling and padding image as input to the network.

    Args:
        img (numpy.ndarray): H x W x C, image read with OpenCV
        input_shape (tuple(int)): input shape of the network for inference
        swap (tuple(int)): new order of axes to transpose the input image

    Returns:
        padded_img (numpy.ndarray): preprocessed image to be fed to the network
        ratio (float): ratio for scaling the image to the input shape
    """
    if len(img.shape) == 3:
        padded_img = np.ones((input_shape[0], input_shape[1], 3), dtype=np.uint8) * 114
    else:
        padded_img = np.ones(input_shape, dtype=np.uint8) * 114
    ratio = min(input_shape[0] / img.shape[0], input_shape[1] / img.shape[1])
    resized_img = cv2.resize(
        img,
        (int(img.shape[1] * ratio), int(img.shape[0] * ratio)),
        interpolation=cv2.INTER_LINEAR,
    ).astype(np.uint8)
    padded_img[: int(img.shape[0] * ratio), : int(img.shape[1] * ratio)] = resized_img
    padded_img = padded_img.transpose(swap)
    padded_img = np.ascontiguousarray(padded_img, dtype=np.float32)
    return padded_img, ratio


def postprocess(outputs, input_shape, ratio):
    """
    Post-processing part of YOLOX for generating final results from outputs of the network.

    Args:
        outputs (tuple(numpy.ndarray)): outputs of the detection heads with onnxruntime session
        input_shape (tuple(int)): input shape of the network for inference
        ratio (float): ratio for scaling the image to the input shape

    Returns:
        dets (numpy.ndarray): n x 6, dets[:,:4] -> boxes, dets[:,4] -> scores, dets[:,5] -> class indices
    """
    outputs = [out.reshape(*out.shape[:2], -1).transpose(0,2,1) for out in outputs]
    outputs = np.concatenate(outputs, axis=1)
    outputs[..., 4:] = sigmoid(outputs[..., 4:])
    predictions = demo_postprocess(outputs, input_shape, p6=False)[0]
    boxes = predictions[:, :4]
    scores = predictions[:, 4:5] * predictions[:, 5:]
    boxes_xyxy = np.ones_like(boxes)
    boxes_xyxy[:, 0] = boxes[:, 0] - boxes[:, 2]/2.
    boxes_xyxy[:, 1] = boxes[:, 1] - boxes[:, 3]/2.
    boxes_xyxy[:, 2] = boxes[:, 0] + boxes[:, 2]/2.
    boxes_xyxy[:, 3] = boxes[:, 1] + boxes[:, 3]/2.
    boxes_xyxy /= ratio
    dets = multiclass_nms(boxes_xyxy, scores, nms_thr=0.45, score_thr=0.1)
    return dets


def sigmoid(x):
    return 1.0 / (1.0 + np.exp(-x))


if __name__ == '__main__':
    args = make_parser().parse_args()
    input_shape = tuple(map(int, args.input_shape.split(',')))
    origin_img = cv2.imread(args.image_path)
    img, ratio = preprocess(origin_img, input_shape)
    if args.ipu:
        providers = ["VitisAIExecutionProvider"]
        provider_options = [{"config_file": args.provider_config}]
    else:
        providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
        provider_options = None
    session = ort.InferenceSession(args.model, providers=providers, provider_options=provider_options)
    ort_inputs = {session.get_inputs()[0].name: img[None, :, :, :]}
    outputs = session.run(None, ort_inputs)
    dets = postprocess(outputs, input_shape, ratio)
    if dets is not None:
        final_boxes, final_scores, final_cls_inds = dets[:, :4], dets[:, 4], dets[:, 5]
        origin_img = vis(origin_img, final_boxes, final_scores, final_cls_inds,
                         conf=args.score_thr, class_names=COCO_CLASSES)
    mkdir(args.output_dir)
    output_path = os.path.join(args.output_dir, os.path.basename(args.image_path))
    cv2.imwrite(output_path, origin_img)