search

zhuhan1236 / mmd-caffe

[DEPRECATED] Please move to https://github.com/thuml/transfer-caffe
Other
9 stars 15 forks source link

single-step debug #1

Closed yhldhit closed 8 years ago

yhldhit commented 8 years ago

When I uncomment the code DEBUG=1 in file "Makefile.config", there is no problem compiling. However, it returns an error when runing, saying: Expr:-CGAL_IA_MUL(-1.1,10.1)!=CGAL_IA_MUL(1.1,10.1) Explanation:Wrong rounding: did you forget the -frounding-math option if you use GCC (or -f-model strict for Intel) I cannot find a solution and can you give me some advices? Thank you.

zhuhan1236 commented 8 years ago

I am sorry that I've never uncommented DEBUG=1. So, I have no idea what actually happened. But according to the ERROR MESSAGE, maybe there's something wrong with CGAL?

yhldhit commented 8 years ago

Thank you @zhuhan1236. It comes out that there is indeed something wrong with CGAL. When I compile the project adding -DCGAL_DISABLE_ROUNDING_MATH_CHECK in variable named NVCCFLAGS in Makefile, I can normally debug this project. But I got a new problem that I am not able to reproduce the results given in the paper. When I deal with the task amazon to webcame, the average in my experiment around 65%, not 68% in the paper. So I will appreciate if you can provide me some suggestions,

zhuhan1236 commented 8 years ago

Thank you for fixing it. Which model did you run? The icml one or the alex_net one?

yhldhit commented 8 years ago

It's my pleasure. And I did experiments on icml model

zhuhan1236 commented 8 years ago

Did you use the sub-sampling dataset or the full dataset of amazon to webcam?

yhldhit commented 8 years ago

I tested on both dataset. And got almost the same result. In addition, a new problem confused me: I firstly borrowed the weights from bvlc. After that, should I fine tune the weights in conv3 to conv5 before training or, instead, directly start training? Thanks. 发自网易邮箱大师 On 2016-06-07 11:03 , zhuhan1236 Wrote: Did you use the sub-sampling dataset or the full dataset of amazon to webcam? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

zhuhan1236 commented 8 years ago

Have you tried to tune the hyper-parameter mmd_lambda? You can directly start training.

yhldhit commented 8 years ago

@zhuhan1236 I used the parameters given in this code and I will try to fine-tune the hyper-parameter in the following experiments thanks to your advice. One more question, the office dataset I downloaded contains only 4110 samples, which is different from 4652 described in the paper. Have you encountered the same problem?

zhuhan1236 commented 8 years ago

Our office dataset contains 4110 samples, too. There may be something wrong when writing the paper. And I'll also tune the parameters after these busy days.

yhldhit commented 8 years ago

@zhuhan1236 Thanks for your time~

yhldhit commented 8 years ago

Hi zhuhan How are you? I discussed with you on Github about the projected mmd-caffe and you did give me some useful advice. I have recently conducted some new experiments based on your implemention. But I have encountered some other problems. One of the problems is that I can reproduce the experimental results stated in the paper without given the detailed experiment setting. Can you offer me more information about this. BTW, do you mind to invite me as your QQ contacts(my QQ number is 1154761720), so that I can contact you easier and we can discuss on the topic of domain adaption in the future. Thanks for your time.

2016-06-14 9:55 GMT+08:00 zhuhan1236 notifications@github.com:

Our office dataset contains 4110 samples, too. There may be something wrong when writing the paper. And I'll also tune the parameters after these busy days.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/zhuhan1236/mmd-caffe/issues/1#issuecomment-225759583, or mute the thread https://github.com/notifications/unsubscribe/AOqPF6n4oXVzgkMB7W5EH2YFQ-njLQECks5qLgopgaJpZM4IqqT_ .

zhuhan1236 commented 8 years ago

OK, I've invited you in QQ

yhldhit commented 8 years ago

Sorry, it seems that I missed your inviting message for some reason. Could you please invite me once again. Thanks~ 发自网易邮箱大师 On 2016-08-30 10:28 , zhuhan1236 Wrote: OK, I've invited you in QQ — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub, or mute the thread.