5.模型评估

首先是导包以及设置context：

from mindspore import contextfrom dataset.voc2012_dataset import create_voc2012_datasetfrom model.yolo import YOLOV3DarkNet53from utils.utils_yolov3 import DetectionEngine, load_yolov3from yolov3_eval_conf import EvalYOLOv3Conf# set contextcontext.set_context(mode=context.GRAPH_MODE,                    device_target="GPU", save_graphs=False)# set_graph_kernel_contextcontext.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")# Set mempool block size for improving memory utilization, which will not take effect in GRAPH_MODEif context.get_context("mode") == context.PYNATIVE_MODE:    context.set_context(mempool_block_size="31GB")# configconfig = 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/"       # 分类检测框的临时结果存放                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目录的位置，注意最后的格式        self.val_path = "VOCdevkit/VOC2012/ImageSets/Main/my_person_val.txt"  # 模型评估xml文件

读数据集：

# datasetvoc2012_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_idxprint("dataset size: ",data_size)print("bath num in 1 epoch: ", config.steps_per_epoch)

定义网络和检测器：

# networknetwork = YOLOV3DarkNet53(is_training=config.data_training, config=config)load_yolov3(network,config.ckpt_path)network.set_train(False)# init detection enginedetection = DetectionEngine(config)

这里我们不需要损失函数，所以只用了网络结构YOLOV3DarkNet53，然后用load_yolov3将前面训练好的模型加载进来。

DetectionEngine类主要处理模型前向传播后的输出，包括置信度阈值筛选、NMS算法去除重叠等等工作，最后计算AP的工作也是它完成，对这些计算感兴趣的朋友可以看 附件\utils\utils_yolov3.py 以及 附件\utils\eval_utils.py。

最后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 pltplt.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")

最后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，得到的结果是AP为0.609，P-R曲线如下：