一、环境搭建

环境信息

ultralytics	8.4.3
pytorch	2.9.1
cann	8.0.1
HDK	24.1.1.3
硬件信息	Atlas300I

driver安装包下载：A300-3010-npu-driver_24.1.1.3_linux-aarch64.run

fireware安装包下载：A300-3010-npu-firmware_7.5.0.9.220.run

cann安装包下载：Ascend-cann-toolkit_8.0.1_linux-aarch64.run

基础环境搭建参考：Qwen2.5-Omni-7B环境搭建及aisbench压测指导

yolo26移植代码开源地址：yolo26-npu

二、模型转换

1、将模型yolo26s.pt导出onnx格式

python3 pth2onnx.py --pt ./models/yolo26s.pt

pth2onnx.py 源代码如下：

import argparse
from ultralytics import YOLO


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--pt', default="./models/yolo26s.pt", help='pt file')
    args = parser.parse_args()

    model = YOLO(args.pt)
    onnx_model = model.export(format="onnx", dynamic=True, simplify=True, opset=11)


if __name__ == '__main__':
    main()

2、atc工具使用说明

文档地址：ATC工具（8.5.0）参数说明

3、将onnx模型转换成om

atc --framework=5 \

    --model=./models/yolo26s.onnx \

    --input_format=NCHW \

    --input_shape='images:1,3,640,640'\

    --output_type=FP32 \

    --output=./models/yolo26s_bs1 \

    --log=info \

    --soc_version=Ascend310

--input_shape: 模型入参尺寸 1*3*640*640

--input_format: 模型入参格式，N 是batchsize，C 图像通道，H 图像高，W 图像宽

--output_type: 模型输出数值是FP32

--soc_version: 推理环境使用的芯片型号

推理设备	soc_version
Atlas300I	Ascend310
Atlas300V	Ascend310P3
Atlas300I Pro	Ascend310P3
Atlas300V Pro	Ascend310P3
Atlas300I Duo	Ascend310P3
Atlas800T A2	Ascend910B3
Atlas800I A2	Ascend910B4

三、模型推理

推理脚本inference.py代码如下：

import os
import json
import argparse
import torch
import numpy as np
from ais_bench.infer.interface import InferSession

from ultralytics.models.yolo.detect import DetectionPredictor
from ultralytics import YOLO

### patch cpu lettebox begin #####
def patch_pre_transform(self, im):
    same_shapes = len({x.shape for x in im}) == 1
    self.model.pt = False
    letterbox = LetterBox(self.imgsz, auto=same_shapes and self.model.pt, stride=self.model.stride)
    return [letterbox(image=x) for x in im]

from ultralytics.data.augment import LetterBox
from ultralytics.engine.predictor import BasePredictor
BasePredictor.pre_transform = patch_pre_transform
### patch cpu lettebox end #####

class OM_Conf():
    def __init__(self):
        self.pt=False
        self.stride = 32
        self.fp16 = False
        self.overrides = {'task': 'detect', 'data': '/usr/local/lib/python3.10/dist-packages/ultralytics/cfg/datasets/coco.yaml', 'imgsz': 640, 'single_cls': False, 'model': './mo
dels/yolo26s_bs1.om', 'conf': 0.25, 'batch': 1, 'save': True, 'mode': 'predict', 'save_txt': True}
        self.names = {0: 'person', 1: 'bicycle', 2: 'car', 3: 'motorcycle', 4: 'airplane', 5: 'bus', 6: 'train', 7: 'truck', 8: 'boat', 9: 'traffic light', 10: 'fire hydrant', 11:
 'stop sign', 12: 'parking meter', 13: 'bench', 14: 'bird', 15: 'cat', 16: 'dog', 17: 'horse', 18: 'sheep', 19: 'cow', 20: 'elephant', 21: 'bear', 22: 'zebra', 23: 'giraffe', 24: 
'backpack', 25: 'umbrella', 26: 'handbag', 27: 'tie', 28: 'suitcase', 29: 'frisbee', 30: 'skis', 31: 'snowboard', 32: 'sports ball', 33: 'kite', 34: 'baseball bat', 35: 'baseball 
glove', 36: 'skateboard', 37: 'surfboard', 38: 'tennis racket', 39: 'bottle', 40: 'wine glass', 41: 'cup', 42: 'fork', 43: 'knife', 44: 'spoon', 45: 'bowl', 46: 'banana', 47: 'app
le', 48: 'sandwich', 49: 'orange', 50: 'broccoli', 51: 'carrot', 52: 'hot dog', 53: 'pizza', 54: 'donut', 55: 'cake', 56: 'chair', 57: 'couch', 58: 'potted plant', 59: 'bed', 60: 
'dining table', 61: 'toilet', 62: 'tv', 63: 'laptop', 64: 'mouse', 65: 'remote', 66: 'keyboard', 67: 'cell phone', 68: 'microwave', 69: 'oven', 70: 'toaster', 71: 'sink', 72: 'ref
rigerator', 73: 'book', 74: 'clock', 75: 'vase', 76: 'scissors', 77: 'teddy bear', 78: 'hair drier', 79: 'toothbrush'}


def save_txt(boxes_data, filepath, pt_result):
    filename = filepath.split("/")[-1]
    res_filepath = pt_result + "/" + filename.replace('.jpg','.txt')

    os.makedirs(os.path.dirname(res_filepath), exist_ok=True)
    with open(res_filepath, "w") as json_fp:
        for box in boxes_data:
            json_dict = dict()
            json_dict["filename"] = filename
            json_dict["x0"] = round(float(box[0]),5)
            json_dict["y0"] = round(float(box[1]),5)
            json_dict["width"] = round((float(box[2]) - json_dict["x0"]),5)
            json_dict["height"] = round((float(box[3]) - json_dict["y0"]),5)
            json_dict["class"] = int(box[5])
            json_dict["score"] = round(float(box[4]),5)
            json_str = json.dumps(json_dict)
            json_fp.write(json_str+'\n')

def pt_detect(input_args):
    model = YOLO(input_args.pt)
    output = model(source=input_args.data, save=True, save_txt=True)  # ndarray
    for i in output:
        print("pt result:", i.boxes.data)
        save_txt(i.boxes.data, i.path, input_args.pt_result)

def om_detect(input_args):
    om_model = InferSession(int(input_args.device_id), input_args.om)

    #data
    om_conf = OM_Conf()
    dp = DetectionPredictor(overrides=om_conf.overrides)
    dp.model = om_conf
    dp.setup_source(input_args.data)
    for dp.batch in dp.dataset:
        paths, im0s, s = dp.batch

        # Preprocess
        im = dp.preprocess(im0s)

        # Inference
        im = np.ascontiguousarray(im).astype(np.float32)  # contiguous
        preds = om_model.infer([im])
        preds_tensor = torch.from_numpy(preds[0])

        # Postprocess
        output = dp.postprocess(preds_tensor, im, im0s)
        print("om result:", output[0].boxes.data)
        save_txt(output[0].boxes.data, output[0].path, input_args.om_result)

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--data', default='./data', help='data path')
    parser.add_argument('--pt', default='./models/yolo26s.pt', help='pt model path')
    parser.add_argument('--om', default='./models/yolo26s_bs1.om', help='om model path')
    parser.add_argument('--pt_result', default='./pt-result', help='pt model result path')
    parser.add_argument('--om_result', default='./om-result', help='om model result path')
    parser.add_argument('--device_id', default='0', help='device id')
    input_args = parser.parse_args()

    pt_detect(input_args)
    om_detect(input_args)

if __name__ == '__main__':
    main()

om模型推理逻辑与ultralytics保持一致：

执行推理脚本并查看pt模型和om模型的推理结果

python3 inference.py

推理结果验证yolo26s.pt 和yolo26s_bs1.om 的推理结果一致

四、计算MAP

1、推理val2017数据集，并保存推理结果

每张图片保存一个推理结果

Map值计算，采用COCO数据集自带MAP计算方法，eval_map.py代码如下：

#!/usr/bin/python

import sys
import os
import json
from pycocotools.coco import COCO 
from pycocotools.cocoeval import COCOeval 

# If necessary, pre-define category and its id
PRE_DEFINE_CATEGORIES = {0: 'person', 1: 'bicycle', 2: 'car', 3: 'motorcycle', 4: 'airplane', 5: 'bus', 6: 'train', 7: 'truck', 8: 'boat', 9: 'traffic light', 10: '
fire hydrant', 11: 'stop sign', 12: 'parking meter', 13: 'bench', 14: 'bird', 15: 'cat', 16: 'dog', 17: 'horse', 18: 'sheep', 19: 'cow', 20: 'elephant', 21: 'bear',
 22: 'zebra', 23: 'giraffe', 24: 'backpack', 25: 'umbrella', 26: 'handbag', 27: 'tie', 28: 'suitcase', 29: 'frisbee', 30: 'skis', 31: 'snowboard', 32: 'sports ball'
, 33: 'kite', 34: 'baseball bat', 35: 'baseball glove', 36: 'skateboard', 37: 'surfboard', 38: 'tennis racket', 39: 'bottle', 40: 'wine glass', 41: 'cup', 42: 'fork
', 43: 'knife', 44: 'spoon', 45: 'bowl', 46: 'banana', 47: 'apple', 48: 'sandwich', 49: 'orange', 50: 'broccoli', 51: 'carrot', 52: 'hot dog', 53: 'pizza', 54: 'don
ut', 55: 'cake', 56: 'chair', 57: 'couch', 58: 'potted plant', 59: 'bed', 60: 'dining table', 61: 'toilet', 62: 'tv', 63: 'laptop', 64: 'mouse', 65: 'remote', 66: '
keyboard', 67: 'cell phone', 68: 'microwave', 69: 'oven', 70: 'toaster', 71: 'sink', 72: 'refrigerator', 73: 'book', 74: 'clock', 75: 'vase', 76: 'scissors', 77: 't
eddy bear', 78: 'hair drier', 79: 'toothbrush'}
CATEGORY_ID = {0: 1, 1: 2, 2: 3, 3: 4, 4: 5, 5: 6, 6: 7, 7: 8, 8: 9, 9: 10, 10: 11, 11: 13, 12: 14, 13: 15, 14: 16, 15: 17, 16: 18, 17: 19, 18: 20, 19: 21, 20: 22, 
21: 23, 22: 24, 23: 25, 24: 27, 25: 28, 26: 31, 27: 32, 28: 33, 29: 34, 30: 35, 31: 36, 32: 37, 33: 38, 34: 39, 35: 40, 36: 41, 37: 42, 38: 43, 39: 44, 40: 46, 41: 
47, 42: 48, 43: 49, 44: 50, 45: 51, 46: 52, 47: 53, 48: 54, 49: 55, 50: 56, 51: 57, 52: 58, 53: 59, 54: 60, 55: 61, 56: 62, 57: 63, 58: 64, 59: 65, 60: 67, 61: 70, 
62: 72, 63: 73, 64: 74, 65: 75, 66: 76, 67: 77, 68: 78, 69: 79, 70: 80, 71: 81, 72: 82, 73: 84, 74: 85, 75: 86, 76: 87, 77: 88, 78: 89, 79: 90}

def get_filename_as_int(filename):
    try:
        filename = filename.replace("\\", "/")
        filename = os.path.splitext(os.path.basename(filename))[0]
        return int(filename)
    except:
        raise ValueError("Filename %s is supposed to be an integer." % (filename))

def get_default_dict():
        return {"image_id": -1, "category_id": -1, "bbox": [], "score": 0}

def xml2json(xml_files):
    json_out = []
    for xml_file in xml_files:
        f = open(xml_file)
        line = f.readline()
        if len(line) == 0:
            #print(xml_file)
            continue
        one = json.loads(line)
        filename = one["filename"]
        ## The filename must be a number
        image_id = get_filename_as_int(filename)

        while line:
            t_dict = get_default_dict()
            t_dict['image_id'] = image_id
            t_dict['category_id'] = CATEGORY_ID[one["class"]]
            t_dict['bbox'] = [one['x0'], one['y0'], one['width'], one['height']]
            t_dict['score'] = one['score'] 
            json_out.append(t_dict)
            line = f.readline()
            if len(line) == 0:
                break
            one = json.loads(line)
        f.close()
    return json_out

def get_img_id(cocoDt_json): 
    ls = [] 
    myset = [] 
    for anno in cocoDt_json: 
      ls.append(anno['image_id']) 
    myset = {}.fromkeys(ls).keys() 
    return myset

'''
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.317
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.562
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.321
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.162
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.343
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.448
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.278
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.438
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.464
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.275
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.497
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.625
'''

def eval_compute(cocoDt_json,cocoGt_file):
    cocoGt = COCO(cocoGt_file)#取得标注集中coco json对象
    imgIds = get_img_id(cocoDt_json) 
    cocoDt = cocoGt.loadRes(cocoDt_json)#取得结果集中image json对象
    imgIds = sorted(imgIds)#按顺序排列coco标注集image_id
    cocoEval = COCOeval(cocoGt, cocoDt, "bbox") 
    cocoEval.params.imgIds = imgIds#参数设置
    cocoEval.evaluate()#评价
    cocoEval.accumulate()#积累
    cocoEval.summarize()#总结

if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser(
        description="Convert Pascal VOC annotation to COCO format."
    )
    parser.add_argument("--xml_dir", default='./om-result', help="Directory path to xml files.", type=str)
    parser.add_argument("--instances_file", default='/data/LLMs/instances_val2017.json', help="COCO format instances_val2017.json.", type=str)
    args = parser.parse_args()

    xml_files= []
    for root, dirs, files in os.walk(args.xml_dir):
        for file in files:
            xml_files.append(os.path.join(root, file))

    # If you want to do train/test split, you can pass a subset of xml files to convert function.
    print("Number of xml files: {}".format(len(xml_files)))
    detJson = xml2json(xml_files)
    eval_compute(detJson, args.instances_file)

注意：yolo26s.pt 推理的标签与val2017数据集的标签需要映射

计算pt模型推理结果的MAP值

python3 eval_map.py --xml_dir pt-result/

计算om模型推理结果的MAP值

python3 eval_map.py --xml_dir om-result/

yolo26模型的MAP值汇总，精度误差控制在千分之一以内

模型	pt 格式 MAP(0.50:0.95)	om 格式 MAP(0.50:0.95)	误差
yolo26n	0.33	0.33	0
yolo26s	0.406	0.405	0.001
yolo26m	0.455	0.454	0.001
yolo26l	0.469	0.469	0
yolo26x	0.499	0.498	0.001