zhizhz commented 1 year ago

老师您好，在我测试您的代码的时候根据readme更换评价指标的时候，是否只需要更改hungarian_matching.py的判断语句： for i in range(prediction_len): for j in range(gt_len):

if EA_metric(p_lines[i], g_lines[j]) >= threshold:

        #if Chamfer_metric(p_lines[i], g_lines[j]) >= threshold:
        if Emd_metric(p_lines[i], g_lines[j]) >= threshold:
            G[i][j] = 1
return G

在我进行不同评价指标测试的时候发现Emd_metric的评价指标的时候会不出结果，而在Chamfer _metric的评价指标的时候会有如下报错： Traceback (most recent call last): File "/home/qiao/app/pycharm-community-2022.2.2/plugins/python-ce/helpers/pydev/pydevd.py", line 1496, in _exec pydev_imports.execfile(file, globals, locals) # execute the script File "/home/qiao/app/pycharm-community-2022.2.2/plugins/python-ce/helpers/pydev/_pydev_imps/_pydev_execfile.py", line 18, in execfile exec(compile(contents+"\n", file, 'exec'), glob, loc) File "/home/qiao/Desktop/deep-hough-transform-master/metric.py", line 83, in emd_score = Emd_metric(l1, l2) File "/home/qiao/Desktop/deep-hough-transform-master/metric.py", line 50, in Emd_metric M = ot.dist(points1, points2, metric='euclidean') File "/home/qiao/anaconda3/lib/python3.7/site-packages/ot/utils.py", line 219, in dist return euclidean_distances(x1, x2, squared=False) File "/home/qiao/anaconda3/lib/python3.7/site-packages/ot/utils.py", line 165, in euclidean_distances nx = get_backend(X, Y) File "/home/qiao/anaconda3/lib/python3.7/site-packages/ot/backend.py", line 82, in get_backend raise ValueError("Unknown type of non implemented backend.") ValueError: Unknown type of non implemented backend. python-BaseException

请问老师，是我更改评价标准的程序步骤没有完全吗还是这个ot.dist函数的原因？

zeakey commented 1 year ago

I guess this is related to the version of ot. OT has recently released a major update that enables multiple backends (e.g. torch and numpy). Your error message tells that it was a 'backend' problem.

You can try to reverse to an older version of ot without multi-backend support.

zhizhz commented 1 year ago

谢谢老师解答，我尝试了

def Emd_metric(l_pred, l_gt, size=(400, 400)): points1 = get_points_coords(l_pred) points2 = get_points_coords(l_gt)

points1 = np.array(points1)
points2 = np.array(points2)
#points1 = torch.cuda.FloatTensor(points1)
#points2 = torch.cuda.FloatTensor(points2)

M = ot.dist(points1, points2, metric='euclidean')

_, log = ot.emd([], [], M, log=True)
cost = log['cost']
return 1 - cost / np.sqrt(size[0] * size[0] + size[1] * size[1])

这样可以在不更改ot版本情况下，解决我上述问题，但是运行python testsel.py文件的时候不会报错但运行时间很长，在训练的时候显示信息如下： Training Loss:0.004758382216095924: 100%|█████| 1300/1300 [08:46<00:00, 2.47it/s] 1%|▎ | 10/1300 [13:04<23:48:51, 66.46s/it] 在验证过程中耗费时间太多，请问老师这是正常的吗，还是说这还是与ot版本有关，需要将ot版本转回旧版本，方便问一下老师您当时POT版本是多少吗

zeakey commented 1 year ago

谢谢老师解答，我尝试了

def Emd_metric(l_pred, l_gt, size=(400, 400)): points1 = get_points_coords(l_pred) points2 = get_points_coords(l_gt)
points1 = np.array(points1)
points2 = np.array(points2)
#points1 = torch.cuda.FloatTensor(points1)
#points2 = torch.cuda.FloatTensor(points2)

M = ot.dist(points1, points2, metric='euclidean')

_, log = ot.emd([], [], M, log=True)
cost = log['cost']
return 1 - cost / np.sqrt(size[0] * size[0] + size[1] * size[1])
这样可以在不更改ot版本情况下，解决我上述问题，但是运行python testsel.py文件的时候不会报错但运行时间很长，在训练的时候显示信息如下： Training Loss:0.004758382216095924: 100%|█████| 1300/1300 [08:46<00:00, 2.47it/s] 1%|▎ | 10/1300 [13:04<23:48:51, 66.46s/it] 在验证过程中耗费时间太多，请问老师这是正常的吗，还是说这还是与ot版本有关，需要将ot版本转回旧版本，方便问一下老师您当时POT版本是多少吗

Any thoughts on this?

Hanqer / deep-hough-transform

更改评价标准的报错问题？ #52

if EA_metric(p_lines[i], g_lines[j]) >= threshold: