zhengpeng7 / sanet-keras Goto Github PK

HELLO, thanks for your open-source.
How about your recurrent performance?
Is this close to the result that paper mentioned?

您好，我现在在做SAnet的复现，然后发现卷积按照same方式，卷积之后特征图会有白色的边，再反卷积的时候，白边会变得越来越明显，最后学到的就是白边的值。下面是最后encode的特征图和decode最后的特征图。

想请问下，可能是哪里出错了。

i try to run generate_datasets, but it came out 'operands could not be broadcast together with shapes (91,811) (91,0) (91,811) '
here is the result

ValueError Traceback (most recent call last)
in
26 k[int(gt[i][1]), int(gt[i][0])] = 1
27
---> 28 k = get_density_map_gaussian( k, gt, adaptive_mode=is_adaptive )
29
30 file_path = img_path.replace('.jpg', '.h5').replace('images', 'ground')

~\SANet\utils.py in get_density_map_gaussian(im, points, adaptive_mode, fixed_value, fixed_values)
65 max(0, p[0]-gaussian_radius):min(density_map.shape[0], p[0]+gaussian_radius+1),
66 max(0, p[1]-gaussian_radius):min(density_map.shape[1], p[1]+gaussian_radius+1)
---> 67 ] += gaussian_map[y_up:y_down, x_left:x_right]
68 # density_map[density_map < 0.0003] = 0
69 density_map = density_map // (np.sum(density_map // num_gt))

ValueError: operands could not be broadcast together with shapes (91,811) (91,0) (91,811)

ValueError: operands could not be broadcast together with shapes (91,811) (91,0) (91,811)

我自己写了一个maee函数用于观测在训练过程中，模型在验证集上面的表现。

但结果发现随着epoch的增加，其值从第一个epoch下降后，一直在增加，希望能得到您的解答。
谢谢！

Hi!
First of all, thank you very much for sharing your code! It's a great work.

My question is if the network could be use with binary images as input.

Thanks a lot!
Best regards.