Closed CGC1031 closed 4 years ago
Please also add a key here, e.g., 'img_moving': img2, at https://github.com/microsoft/Recursive-Cascaded-Networks/blob/0709490fe010e4acc6357990d1d871c170e3ef31/network/recursive_cascaded_networks.py#L176
Thank you for your reply. I have achieved it successfully. Thanks again.
您好,请问是把eval.py中keys = ['pt_mask', 'landmark_dists', 'jaccs', 'dices', 'jacobian_det', ]替换为keys = ['real_flow', 'image_fixed', 'warped_moving']就可以了吗?还需要修改其他地方吗?
期待您的回复。
尊敬的赵博士您好! 我研究您“Recursive cascaded networks for unsupervised medical image registration”三年了,并在您的工作基础上做了一些研究作为自己的毕设。最近毕业论文的审核老师提出一个疑问,我自己也难以解答,所以冒昧致信向您求教,还望赵博士帮我解解疑惑。 下图中仿射子网络Affine和形变子网络Dense您分别用两个损失函数来进行约束,我想请问一个网络训练为什么会有两个损失函数呢? 假设一:您是不是先用仿射配准子网络的损失函数loss1训练好仿射配准网络的12个形变参数,然后将其固定;再用形变配准子网络的损失函数loss2去训练形变网络的参数。总之就是训练时只有一个loss在起作用,Affine是一个部分,后面的所有形变网络是一个部分,分开训练的。 假设二:loss1指导Affine训练,loss2指导Dense训练,但是二者是同步的,一直训练到结束。整个训练过程中,两个loss同时工作,同时指导Affine和Dense网络参数地不断优化(我也不知道卷积神经网络是否可以同时被两个损失函数所指导?) 非常期待您回复,感谢您!
The whole network can optimize multiple objectives simultaneously.
Thank you for your patient reply before.Your advice is helpful to me. I have successfully get the real_flow, warped_moving and image_fixed pictures. Through add keys = ['real_flow', 'image_fixed', 'warped_moving'] . For the purpose of getting moving image, following the previous method, I attempted to add the following key words to keys. ['image_moving'] / ['img_moving'] / ['moving'] . But they are all failed.
Prompt us for keyword error['img_moving'].
1、Can you tell me what the key words correspond to the moving image?
2、Or how do you get moving images?Can you give me some ideas?
Looking forward to your reply!