search

nitishsrivastava / deepnet

Implementation of some deep learning algorithms.
BSD 3-Clause "New" or "Revised" License
893 stars 438 forks source link

Anyone used eigenmat? #89

Open moyuanhuang opened 8 years ago

moyuanhuang commented 8 years ago

Hi Nitish, I choose too use your eigenmat library since my laptop's GPU isn't compatible with cudamat. Then I've discovered several bugs while using the eigenmat.I fixed trivial ones which I'll list below. 1./eigenmat/Makefile: all LIB_NAME should be libeigenmat instead of lineigenmat 2./deepnet/extract_rbm_representation.py in main(): line 55 & line77: LockGPU() and FreeGPU() should no longer be called since cudamat is not used.

By modifying the code, I successfully run the image modality of the multimodal-dbn(which implemented the Gaussian Boltzmann Machine). I even used my own dataset extracted from music tracks. Everything goes perfect until run into the text modality. It crushed just before the training begin. I paste the error message here and is kind of confusing:

Training first layer text RBM. Failed to import cudamat. Using eigenmat. No GPU will be used. Train Step: 0Assertion failed: (aLhs.rows() == aRhs.rows() && aLhs.cols() == aRhs.cols()), function CwiseBinaryOp, file /usr/local/include/eigen3/Eigen/src/Core/CwiseBinaryOp.h, line 132. ./runall_dbn.sh: line 78: 1210 Abort trap: 6 python ${trainer} models/text_rbm1.pbtxt trainers/dbn/train_CD_text_layer1.pbtxt eval.pbtxt

I once thought it could be my dataset's problem, so then I used the flickr dataset to run the text modality training, and it provides me with the same error messages.

tingkuei commented 8 years ago

Is this problem solved?? I got the same error messages while running the first text modality.