ultralytics库导出onnx模型，模型失去预测能力

Question

今天用ultralytics库把一个yolov8n模型导出成onnx格式,验证了一下,发现导出之后的模型相比.pt模型预测完全不准。导出代码如下：

from ultralytics import YOLO
import torch

# Load a model
model = YOLO('yolov8n.yaml')  # build a new model from YAML
model = YOLO('yolov8n.pt')  # load a pretrained model (recommended for training)
model = YOLO('yolov8n.yaml').load('yolov8n.pt')  # build from YAML and transfer weights

# Train the model
# results = model.train(data='coco128_.yaml', epochs=100, imgsz=640)
# if __name__ == "__main__":
#     torch.onnx.export(model, torch.rand(1, 3, 640, 640), "yolov8n.onnx",
#                       dynamic_axes={'input.1' : {0 : 'batch_size'}, 'output' : {0 : 'batch_size'}})

model.export(format="onnx")

随后验证这个onnx模型的有效性,代码如下：

import onnxruntime as ort
from PIL import Image
import numpy as np
from ultralytics.utils import ops
import torch


# 加载ONNX模型
def load_model(model_path):
    sess = ort.InferenceSession(model_path)
    return sess


# 对图片进行预处理
def preprocess_image(image_path, target_size=(640, 640), if_mask=True, if_show=False):
    # 加载图片并调整大小,适用于1920*1080的原图片
    img = Image.open(image_path)
    size = img.size
    if size[0] > size[1] and if_mask:
        img = img.resize((target_size[0], target_size[1]*size[1]//size[0]))
    else:
        img = img.resize((target_size[0], target_size[1]))
    # 将图片转换为numpy数组
    img_array = np.array(img)
    img_array = img_array
    img_array = np.transpose(img_array, (2, 0, 1))
    if if_mask:
        ch, h, w = img_array.shape
        mask_ch = ch
        mask_h = (w - h)//2
        mask_w = w
        mask = np.zeros((mask_ch, mask_h, mask_w), dtype=np.float32)
        img_array = np.concatenate((mask, img_array, mask), axis=1)
    if if_show:
        img_array_ = img_array.copy()
        img_array_ = np.transpose(img_array_, (1, 2, 0))
        image = Image.fromarray(img_array_.astype(np.uint8))
        image.show()
    img_array = np.expand_dims(img_array, axis=0)
    # 转换数据类型
    img_array = img_array.astype(np.float32)
    return img_array


# 对图片进行预测
def predict(model, image_path):
    # 加载图片并预处理
    img_array = preprocess_image(image_path)
    # 获取模型输入和输出名称
    input_name = model.get_inputs()[0].name
    output_name = model.get_outputs()[0].name
    # 进行预测
    pred_onnx = model.run([output_name], {input_name: img_array})[0]
    return pred_onnx


# 主函数
def main(model_path, image_path):
    # 加载模型
    model = load_model(model_path)
    # 对图片进行预测
    predictions = predict(model, image_path)
    predictions = torch.tensor(predictions)
    preds = ops.non_max_suppression(predictions, conf_thres=0.5, iou_thres=0.25, nc=80, max_det=10)
    predictions_ = np.array(preds)
    print("Predictions:", predictions_)


# 设置模型路径和图片路径
model_path = "yolov8n.onnx"
image_path = "../video_process/processed_photos/frame_0100.jpg"
# 运行主函数
main(model_path, image_path)

这个是我用来预测的样图：

这个是用ultralytics下载的.pt文件的预测结果：

看起来完全正常。
但是,我用ultralytics把它导出成.onnx文件之后,我尝试了两种处理图片的方法,都处理成了（1,3,640,640）的张量

但是,预测的结果大差不差都是这样：

我是选的前10个置信度最大的结果,但是这个里面都是第45,49类,应该有至少两个人（第0类的,而且这个框的宽度连一个像素都没有,很明显这个模型是失效的。如何解决？

Answer 1

博主你好，我根据你的问题，我帮你重新写了个代码，您看下准确率如何，有没有解决你的问题。

# Ultralytics YOLO 🚀, AGPL-3.0 license

import argparse

import cv2
import numpy as np
import onnxruntime as ort
import torch

# from ultralytics.utils import ASSETS, yaml_load
# from ultralytics.utils.checks import check_requirements, check_yaml
from names import names

class YOLOv8:
    """YOLOv8 object detection model class for handling inference and visualization."""

    def __init__(self, onnx_model, input_image, confidence_thres, iou_thres, if_show=False):
        """
        Initializes an instance of the YOLOv8 class.

        Args:
            onnx_model: Path to the ONNX model.
            input_image: Path to the input image.
            confidence_thres: Confidence threshold for filtering detections.
            iou_thres: IoU (Intersection over Union) threshold for non-maximum suppression.
        """
        self.if_show = if_show
        self.onnx_model = onnx_model
        self.input_image = input_image
        self.confidence_thres = confidence_thres
        self.iou_thres = iou_thres

        # Load the class names from the COCO dataset
        self.classes = names

        # Generate a color palette for the classes
        self.color_palette = np.random.uniform(0, 255, size=(len(self.classes), 3))

    def draw_detections(self, img, box, score, class_id):
        """
        Draws bounding boxes and labels on the input image based on the detected objects.

        Args:
            img: The input image to draw detections on.
            box: Detected bounding box.
            score: Corresponding detection score.
            class_id: Class ID for the detected object.

        Returns:
            None
        """

        # Extract the coordinates of the bounding box
        x1, y1, w, h = box

        # Retrieve the color for the class ID
        color = self.color_palette[class_id]

        # Draw the bounding box on the image
        cv2.rectangle(img, (int(x1), int(y1)), (int(x1 + w), int(y1 + h)), color, 2)

        # Create the label text with class name and score
        label = f"{self.classes[class_id]}: {score:.2f}"

        # Calculate the dimensions of the label text
        (label_width, label_height), _ = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)

        # Calculate the position of the label text
        label_x = x1
        label_y = y1 - 10 if y1 - 10 > label_height else y1 + 10

        # Draw a filled rectangle as the background for the label text
        cv2.rectangle(
            img, (label_x, label_y - label_height), (label_x + label_width, label_y + label_height), color, cv2.FILLED
        )

        # Draw the label text on the image
        cv2.putText(img, label, (label_x, label_y), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 1, cv2.LINE_AA)

    def preprocess(self):
        """
        Preprocesses the input image before performing inference.

        Returns:
            image_data: Preprocessed image data ready for inference.
        """
        # Read the input image using OpenCV
        self.img = cv2.imread(self.input_image)

        # Get the height and width of the input image
        self.img_height, self.img_width = self.img.shape[:2]

        # Convert the image color space from BGR to RGB
        img = cv2.cvtColor(self.img, cv2.COLOR_BGR2RGB)

        # Resize the image to match the input shape
        img = cv2.resize(img, (self.input_width, self.input_height))

        # Normalize the image data by dividing it by 255.0
        image_data = np.array(img) / 255.0

        # Transpose the image to have the channel dimension as the first dimension
        image_data = np.transpose(image_data, (2, 0, 1))  # Channel first

        # Expand the dimensions of the image data to match the expected input shape
        image_data = np.expand_dims(image_data, axis=0).astype(np.float32)

        # Return the preprocessed image data
        return image_data

    def postprocess(self, input_image, output):
        """
        Performs post-processing on the model's output to extract bounding boxes, scores, and class IDs.

        Args:
            input_image (numpy.ndarray): The input image.
            output (numpy.ndarray): The output of the model.

        Returns:
            numpy.ndarray: The input image with detections drawn on it.
        """

        # Transpose and squeeze the output to match the expected shape
        outputs = np.transpose(np.squeeze(output[0]))

        # Get the number of rows in the outputs array
        rows = outputs.shape[0]

        # Lists to store the bounding boxes, scores, and class IDs of the detections
        boxes = []
        scores = []
        class_ids = []

        # Calculate the scaling factors for the bounding box coordinates
        x_factor = self.img_width / self.input_width
        y_factor = self.img_height / self.input_height

        # Iterate over each row in the outputs array
        for i in range(rows):
            # Extract the class scores from the current row
            classes_scores = outputs[i][4:]

            # Find the maximum score among the class scores
            max_score = np.amax(classes_scores)

            # If the maximum score is above the confidence threshold
            if max_score >= self.confidence_thres:
                # Get the class ID with the highest score
                class_id = np.argmax(classes_scores)

                # Extract the bounding box coordinates from the current row
                x, y, w, h = outputs[i][0], outputs[i][1], outputs[i][2], outputs[i][3]

                # Calculate the scaled coordinates of the bounding box
                left = int((x - w / 2) * x_factor)
                top = int((y - h / 2) * y_factor)
                width = int(w * x_factor)
                height = int(h * y_factor)

                # Add the class ID, score, and box coordinates to the respective lists
                class_ids.append(class_id)
                scores.append(max_score)
                boxes.append([left, top, width, height])

        # Apply non-maximum suppression to filter out overlapping bounding boxes
        indices = cv2.dnn.NMSBoxes(boxes, scores, self.confidence_thres, self.iou_thres)

        # Iterate over the selected indices after non-maximum suppression
        detected = []
        for i in indices:
            # Get the box, score, and class ID corresponding to the index
            box = boxes[i]
            score = scores[i]
            class_id = class_ids[i]
            detected.append({'box': box, 'score': score, 'class_id': class_id})
            # Draw the detection on the input image
            if self.if_show:
                self.draw_detections(input_image, box, score, class_id)

        # Return the modified input image
        if self.if_show:
            return input_image
        else:
            return detected

    def main(self):
        """
        Performs inference using an ONNX model and returns the output image with drawn detections.

        Returns:
            output_img: The output image with drawn detections.
        """
        # Create an inference session using the ONNX model and specify execution providers
        provider1 = ["CUDAExecutionProvider", "CPUExecutionProvider"]
        provider2 = ["CPUExecutionProvider"]
        session = ort.InferenceSession(self.onnx_model, providers=provider2)

        # Get the model inputs
        model_inputs = session.get_inputs()

        # Store the shape of the input for later use
        input_shape = model_inputs[0].shape
        self.input_width = input_shape[2]
        self.input_height = input_shape[3]

        # Preprocess the image data
        img_data = self.preprocess()

        # Run inference using the preprocessed image data
        outputs = session.run(None, {model_inputs[0].name: img_data})

        # Perform post-processing on the outputs to obtain output image.
        return self.postprocess(self.img, outputs)  # output image


if __name__ == "__main__":
    # Create an argument parser to handle command-line arguments
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", type=str, default="yolov8n.onnx", help="Input your ONNX model.")
    parser.add_argument("--img", type=str, default="../video_process/processed_photos/frame_0100.jpg", help="Path to input image.")
    parser.add_argument("--conf-thres", type=float, default=0.5, help="Confidence threshold")
    parser.add_argument("--iou-thres", type=float, default=0.5, help="NMS IoU threshold")
    parser.add_argument("--if-show", type=bool, default=False, help="If show the output image.")
    args = parser.parse_args()

    # Check the requirements and select the appropriate backend (CPU or GPU)
    # check_requirements("onnxruntime-gpu" if torch.cuda.is_available() else "onnxruntime")

    # Create an instance of the YOLOv8 class with the specified arguments
    import time
    time_start = time.time()
    detection = YOLOv8(args.model, args.img, args.conf_thres, args.iou_thres, args.if_show)

    # Perform object detection and obtain the output image
    output = detection.main()

    if detection.if_show:
        # Display the output image in a window
        cv2.namedWindow("Output", cv2.WINDOW_NORMAL)
        cv2.imshow("Output", output)

        # Wait for a key press to exit
        cv2.waitKey(0)
    else:
        print(output)
    time_end = time.time()
    print("time cost: ", time_end - time_start)