0

I'm new with Caffe and all the ML stuff but I've been working with LeNet using LMDB files without any problem. Now, making a small test over my new Net (FCN32 semantic segmentation) it seems that always get stuck forever trying to open the file. I try with CPU and GPU. Any comment about it is welcome.

 I0614 12:34:29.158532 27812 net.cpp:100] Creating Layer mnist
 I0614 12:34:29.158545 27812 net.cpp:408] mnist -> data
 I0614 12:34:29.158573 27812 net.cpp:408] mnist -> label
 I0614 12:34:29.169967 27818 db_lmdb.cpp:35] Opened lmdb 
 /user/myname/LMDB_file
msanford
  • 11,803
  • 11
  • 66
  • 93
byaeggy
  • 1
  • 1
    is it possible both train and test phase of your net are accessing the same lmdb? it seems like there's a lock on your dataset – Shai Jun 14 '17 at 19:54
  • Hi Shai!, thanks for your answer but nope, I'm working with different lmdb for test and train and both aren't lock. I was thinking in something related to the kernel size (lack of memory?) or try it with a even smaller net (1 layer maybe?). – byaeggy Jun 14 '17 at 23:31
  • doesn't sound like memory issue. can you please post the prototxt part defining data layers? – Shai Jun 15 '17 at 04:14
  • please do not post such comments: it's unreadable. Edit your question instead – Shai Jun 15 '17 at 22:16
  • what about TEST phase input layer? – Shai Jun 15 '17 at 22:16
  • Thanks Shai!, my stuff is here: https://github.com/byaeggy/SSNet – byaeggy Jun 15 '17 at 22:20
  • your train and validation nets are the same. they both read from the SAME lmdb. – Shai Jun 15 '17 at 22:24
  • I'm giving different files, "pid_LMDB1X" for validation and "energy_LMDB1X" for train, do I need to have a different "net structure" for validation and train?. – byaeggy Jun 15 '17 at 22:28
  • Another thing, in my solver sometimes I set train and validation and sometimes I just turn off the validation, and in both cases I get the same issue. – byaeggy Jun 15 '17 at 22:31

0 Answers0