llkn-2

@shibefore @MhLiao 请问一下二位，这个问题解决了吗？我在CTW1500数据集上训练，在res18上面F-score只有67%；将thresh设置成0.05后，F-score提升到了74%；将epsilon 改成0.02，F-score反而只有58%......

The computer memory is 240G.

Yes! When `A.shape` is `(128, 64, 3, 3),` there is no problem. However, when the shape increase to `(256, 64, 3, 3)`， the segmentation fault happens. And I also try...

I changed to `pytorch` backen, and the problem was solved when the shape was `(256, 64, 3, 3)`. But when I continued increasing the size to `(512, 256, 3, 3)`,...

Oh！ I have found where the problem is. When I used `init='random'`, the code just worked well. The default value is `init='svd'`.

OK, thanks for your reply~~It's encouraging that this package is still keeping maintenance.