test.py

import tqdm
import numpy as np

import torch
from torch.autograd import Variable

from utils.datasets import *
from utils.utils import *



def evaluate(model, path, iou_thres, conf_thres, nms_thres, img_size, batch_size):
    model.eval()  # 验证模式

    # 加载验证集
    dataset = ListDataset(path, img_size=img_size, augment=False, multiscale=False)
    dataloader = torch.utils.data.DataLoader(
        dataset, batch_size=batch_size, shuffle=False, num_workers=1, collate_fn=dataset.collate_fn
    )

    # GPU
    Tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.FloatTensor

    labels = []
    sample_metrics = []
    for batch_i, (_, imgs, targets) in enumerate(tqdm.tqdm(dataloader, desc="Detecting objects")):

        # Extract labels
        labels += targets[:, 1].tolist()
        # Rescale target
        targets[:, 2:] = xywh2xyxy(targets[:, 2:])
        targets[:, 2:] *= img_size

        imgs = Variable(imgs.type(Tensor), requires_grad=False)

        with torch.no_grad():
            outputs = model(imgs)
            outputs = non_max_suppression(outputs, conf_thres=conf_thres, nms_thres=nms_thres)

        sample_metrics += get_batch_statistics(outputs, targets, iou_threshold=iou_thres)

    # Concatenate sample statistics
    true_positives, pred_scores, pred_labels = [np.concatenate(x, 0) for x in list(zip(*sample_metrics))]
    precision, recall, AP, f1, ap_class = ap_per_class(true_positives, pred_scores, pred_labels, labels)

    return precision, recall, AP, f1, ap_class