renmin1991/dyamic-graph-representation

Official Dynamic Graph Representation PyTorch implement for iris/face recognition

dyamic-graph-representation's Issues

The loss turned nan

Thank u for your excellent work!

I was training on my dataset and the loss suddenly turned nan at around step 5000. I did the same training on the same dataset a few days ago and it went alright. The only thing I have changed is the total steps in the config_train_singlescale.py.

I think there is something wrong with the loss function. I tried to solve this but couldn't.

Label.txt

您好，您的这篇文章在使用GAT进行虹膜特征匹配上取得了良好的效果，我想复现您的这篇文章，关于您使用的CASIA-Iris数据集，所使用的标注包括哪些内容，如果可以的话能否分享一下，谢谢，[email protected]

Training issue - model learning nothing!

Hi, nice work.
I tried training the network for 9000 classes of iris images normalised to 512*64 using rubber sheet algo, with total number of 25k images. I commented the resize transform. I am doing single scale first!!

The t_loss started with a value of 5 and after 5k/10k iterations the loss was just reduced to 4.6

Also there was no significant difference in matching scores and non matching scores even on training set.

Can you please help me? Let me know if you need any additional information.

关于深层网络的迁移问题

你好~

我看到在README里有人脸识别的结果，包括resnet50和resnet101，但是code里好像是一个5层CNN。

所以想请教一下对于resnet的深层网络，怎么将DGR迁移过去，需要做出哪些改变？

谢谢~

模型大小优化

作者你好，我使用了提供的与训练模型进行虹膜匹配，效果表示的还不错。但是该单尺度模型大小30M，如果想要减轻模型，有哪些优化可以做么？

ND虹膜数据集

您好，可以分享一下ND的虹膜数据集嘛，申请了之后一直获取不到

About graph_pair_sim function

Hello~An excellent work!

When reading your code, i find that the implement of function "graph_pair_sim" is different from loss.py and match_fn.py

Code in match_fn.py

dis_loc = (loc_dis_mat * weight_mat).mean()

Code in loss.py

dis_loc = (loc_dis_mat * weight_mat).sum()

I think cause goal of script is different, so when matching it is mean(), and when calculating loss it is sum(), all right？

May be a simple question, but what are the labels in the labels.txt referring to

After the filename in the txt file, does the label given represent a left or right eye? Or does it mean something else entirely?

test文件只能测试single的模型 multiscale的模型怎么测试