Evaluation result

[asdjia]

Hi,

I appreciate your code. I was wondering how to get the evaluation results for each testing dataset.

I appreciate your help.

Thanks

[CauchyComplete]

Hi,
I'm uploading test_val_v2 function.

def testval_v2(testloader, model):
    avg_acc = AverageMeter()
    avg_p_acc = AverageMeter()
    avg_mIoU = AverageMeter()
    avg_p_mIoU = AverageMeter()
    avg_F1 = AverageMeter()
    avg_p_F1 = AverageMeter()
    avg_AP = AverageMeter()
    avg_p_AP = AverageMeter()

    with torch.no_grad():
        for index, (image, label, qtable) in enumerate(tqdm(testloader)):
            size = label.size()
            image = image.cuda()
            label = label.long().cuda()
            model.eval()
            _, pred = model(image, label, qtable)

            pred = torch.squeeze(pred, 0)
            pred = F.softmax(pred, dim=0)[1]
            pred = pred.unsqueeze(0).unsqueeze(0)

            if pred.size()[-2] != size[-2] or pred.size()[-1] != size[-1]:
                pred = F.upsample(pred, (size[-2], size[-1]), mode='bilinear', align_corners=False)

            pred_r = pred.cpu().numpy().squeeze(axis=0).squeeze(axis=0).ravel()
            label_r = label.cpu().numpy().squeeze(axis=0).ravel()
            pred_r = pred_r[label_r != -1]
            label_r = label_r[label_r != -1]
            bin_pred = (pred_r >= 0.5).astype(np.float)
            correct = (bin_pred == label_r).astype(np.float)
            incorrect = (bin_pred != label_r).astype(np.float)
            TP = np.count_nonzero(correct[label_r==1])
            TN = np.count_nonzero(correct[label_r==0])
            FP = np.count_nonzero(incorrect[label_r==0])
            FN = np.count_nonzero(incorrect[label_r==1])

            mean_IoU = 0.5 * (TP / np.maximum(1.0, TP + FP + FN)) + 0.5 * (TN / np.maximum(1.0, FP + TN + FN))
            avg_mIoU.update(mean_IoU)
            p_mIoU = 0.5 * (FN / np.maximum(1.0, FN + TP + TN)) + 0.5 * (FP / np.maximum(1.0, FP + TP + TN))
            avg_p_mIoU.update(np.maximum(mean_IoU, p_mIoU))
            acc = (TP+TN)/(TP+TN+FP+FN)
            avg_acc.update(acc)
            p_acc = np.maximum(acc, (FP+FN)/(TP+TN+FP+FN))
            avg_p_acc.update(p_acc)
            F1 = (2*TP) / np.maximum(1.0, 2*TP+FN+FP)
            avg_F1.update(F1)
            p_f1 = (2*FN) / np.maximum(1.0, 2*FN+TP+TN)
            avg_p_F1.update(np.maximum(F1, p_f1))

            AP = average_precision_score(label_r, pred_r)
            avg_AP.update(AP)
            p_AP = average_precision_score(label_r, 1-pred_r)
            avg_p_AP.update(np.maximum(AP, p_AP))


            del pred
            torch.cuda.empty_cache()
            gc.collect()

    results = {'avg_p_acc': avg_p_acc.average(),
               'avg_p_mIoU': avg_p_mIoU.average(),
               'avg_p_F1': avg_p_F1.average(),
               'avg_p_AP': avg_p_AP.average(),
               }

    return results

[asdjia]

Thanks for your reply. I am trying to test using trained model. But it was a struggle for me to process the test dataset. I wonder if you can provide some examples about that.

I appreciate your help!