关于人工智能:MindSporeYOLOv3人体目标检测模型实现四

5. 模型评估首先是导包以及设置 context：from mindspore import context

from dataset.voc2012_dataset import create_voc2012_dataset
from model.yolo import YOLOV3DarkNet53
from utils.utils_yolov3 import DetectionEngine, load_yolov3
from yolov3_eval_conf import EvalYOLOv3Conf

context.set_context(mode=context.GRAPH_MODE,

                device_target="GPU", save_graphs=False)

context.set_context(enable_graph_kernel=True)
context.set_context(graph_kernel_flags=”–enable_parallel_fusion “

                                   "--enable_trans_op_optimize"
                                   "--disable_cluster_ops=ReduceMax,Reshape"
                                   "--enable_expand_ops=Conv2D")

if context.get_context(“mode”) == context.PYNATIVE_MODE:

context.set_context(mempool_block_size="31GB")

config = EvalYOLOv3Conf() 评估模型时的配置如下（也可见附件 \yolov3_eval_conf.py）：class EvalYOLOv3Conf():

def __init__(self):
    
    # ---- dataset ----
    
    self.data_path = "VOCdevkit/VOC2012/"
    self.data_usage = "my_person_val"       # 评估数据集
    self.data_training = False
    
    self.num_classes = 1
    
    self.class_to_idx = {}
    
    self.anchor_scales = [[15, 38],
                          [34, 86],
                          [84, 127],
                            
                          [51, 192],
                          [91, 257],
                          [173, 195],
                            
                          [142, 319],
                          [221, 339],
                          [351, 365]]
    
    self.batch_size = 48                      # 评估时的 batch
    self.max_box = 32
    
    # test
    self.test_img_shape = [416, 416]          # 图片缩放大小
    
    # ---- Model ----
    
    self.out_channel = 3 * (5 + self.num_classes)
    
    self.keep_detect = True
    
    self.ckpt_path = "./train_ckpt/yolov3-???.ckpt" # 评估的模型
    
    # ---- Detect ----
    
    self.nms_thresh = 0.5                     # nms 算法去重的 IoU 阈值
    self.eval_ignore_threshold = 0.001        # 检测置信度阈值
    
    self.detcte_result_dir = "det_res/"       # 分类检测框的长期后果寄存
    
    self.image_id_idx = []                    # xml 编号的列表，与图片序号对应，eval_yolov3.py 中设置
    
    self.anno_path = "VOCdevkit/VOC2012/Annotations/{}.xml"               # Annotations 目录的地位，留神最初的格局
    self.val_path = "VOCdevkit/VOC2012/ImageSets/Main/my_person_val.txt"  # 模型评估 xml 文件读数据集：# dataset

voc2012_dat, data_size = create_voc2012_dataset(config, 2)
config.steps_per_epoch = int(data_size / config.batch_size)
image_id_idx = {}
with open(config.val_path) as f:

lines = f.readlines()
for i, line in enumerate(lines):
    image_id_idx[i] = line.strip()

config.image_id_idx = image_id_idx
print(“dataset size: “,data_size)
print(“bath num in 1 epoch: “, config.steps_per_epoch) 定义网络和检测器：# network
network = YOLOV3DarkNet53(is_training=config.data_training, config=config)
load_yolov3(network,config.ckpt_path)
network.set_train(False)

detection = DetectionEngine(config) 这里咱们不须要损失函数，所以只用了网络结构 YOLOV3DarkNet53，而后用 load_yolov3 将后面训练好的模型加载进来。DetectionEngine 类次要解决模型前向流传后的输入，包含置信度阈值筛选、NMS 算法去除重叠等等工作，最初计算 AP 的工作也是它实现，对这些计算感兴趣的敌人能够看附件 \utils\utils_yolov3.py 以及附件 \utils\eval_utils.py。最初对模型进行测试评估：print(‘Start inference….’)
for i, data in enumerate(voc2012_dat.create_dict_iterator(num_epochs=1)):

image = data["image"]

image_shape = data["image_shape"]
image_id = data["img_id"]

prediction = network(image)
output_big, output_me, output_small = prediction
output_big = output_big.asnumpy()
output_me = output_me.asnumpy()
output_small = output_small.asnumpy()
image_id = image_id.asnumpy()
image_shape = image_shape.asnumpy()

detection.detect([output_small, output_me, output_big], config.batch_size, image_shape, image_id)
if i % 2 == 0:
    print('Processing... {:.2f}%'.format(i * config.batch_size / data_size * 100))

print(“Finish”)

print(‘Calculating mAP…’)
detection.do_nms_for_results()
result_file_path = detection.write_result()
print(‘result file path: ‘, result_file_path)

detection.get_eval_result()

for k, t in detection.eval_res.items():

print(k, "AP :", t['ap'])

import matplotlib.pyplot as plt

plt.title(“P-R curve”,fontsize=14)
plt.xlabel(“recall”, fontsize=14)
plt.ylabel(“precision”, fontsize=14)
res = detection.eval_res[‘person’]
plt.plot(res[‘prec’], res[‘rec’])

plt.savefig(“p_r.png”) 最初我设置检测置信度的阈值为 0.01，nms 算法中的 IoU 阈值为 0.5 时，失去的后果是 AP 为 0.609，P- R 曲线如下：

关于人工智能:MindSporeYOLOv3人体目标检测模型实现四

set context

set_graph_kernel_context

Set mempool block size for improving memory utilization, which will not take effect in GRAPH_MODE

config

init detection engine