Closed ZhuangNi closed 7 years ago
Just as I said in #2 , the implementation is not elegant and I modified a lot of original caffe layers like data_layer, softmax_loss_layer and so on. So, the correctness of mmd_layer depends on the whole mmd-caffe project. If you want to run this code, you should clone and compile the complete project. And I'am going to publish a simplified and independent 'mmd_layer' in a few days.
Thank you very much! I just use the latest MMD implemention ,and it could work.
You mean the latest master branch? It does work but a bit inconvenient for users...
I added the MMD Loss to the latest caffe version and use it to train my own data, but the accuracy is just 0.5675. It just start train.
Anyway, thank you very much! I will find the other bug of the program .
Please move to transfer-caffe
OK! You are so nice! Thank you very much!
That's fine.
Hi,zhuhan.When I added your implemention of MMD to the lastest caffe version,I met the following error:
*** Aborted at 1478502296 (unix time) try "date -d @1478502296" if you are using GNU date *** PC: @ 0x7f645ee615a6 caffe::Blob<>::Reshape() *** SIGSEGV (@0x0) received by PID 21540 (TID 0x7f645f8b9a40) from PID 0; stack trace: *** @ 0x7f645d7d5cb0 (unknown) @ 0x7f645ee615a6 caffe::Blob<>::Reshape() @ 0x7f645edf645a caffe::Net<>::ForwardFromTo() @ 0x7f645edf67f7 caffe::Net<>::Forward() @ 0x7f645ee52807 caffe::Solver<>::Step() @ 0x7f645ee530c9 caffe::Solver<>::Solve() @ 0x4094ab train() @ 0x4068ac main @ 0x7f645d7c0f45 (unknown) @ 0x40711b (unknown) @ 0x0 (unknown) Segmentation fault (core dumped)
It works well before the program runif(now_iter_>=iter_of_epoch) ,..,blablabla
command. Could you tell me what happens? Thanks in advance.