0

Hi guys I create a HDF5 dataset from rgb images using matlab with examples from caffe. Before i create the dataset I convert my images for gray scale.

The dataset is created without problems, but when i start training I got this error:

I0501 08:54:43.176820 14655 net.cpp:106] Creating Layer conv3
I0501 08:54:43.176826 14655 net.cpp:454] conv3 <- conv2
I0501 08:54:43.176831 14655 net.cpp:411] conv3 -> conv3
I0501 08:54:43.177955 14655 net.cpp:150] Setting up conv3
I0501 08:54:43.177970 14655 net.cpp:157] Top shape: 5 16 114 114 (1039680)
I0501 08:54:43.177975 14655 net.cpp:165] Memory required for data: 51416320
I0501 08:54:43.177983 14655 layer_factory.hpp:76] Creating layer loss
I0501 08:54:43.177991 14655 net.cpp:106] Creating Layer loss
I0501 08:54:43.177995 14655 net.cpp:454] loss <- conv3
I0501 08:54:43.178000 14655 net.cpp:454] loss <- label
I0501 08:54:43.178006 14655 net.cpp:411] loss -> loss
F0501 08:54:43.178031 14655 euclidean_loss_layer.cpp:12] Check failed: bottom[0]->count(1) == bottom[1]->count(1) (207936 vs. 1024) Input
*** Check failure stack trace: ***

I change the bottom of euclidean layer setting both on conv3 and works, but this is wrong right?

Btw i using this network for image reconstruction and my data is a 4D array composed of data=[12342, 1, 32, 32] and labels=[12342,1,32,32]. I change the dimensions of label array but does not work.

Someone have any idea?

Thanks.

Antonio Paes
  • 11
  • 6
  • This is not duplicate i read the link you suggest the problem that link is on constructions of data layer which says of something is wrong with the dataset, my problem is on the output of the network. So please remove the DUPLICATE and read the question with attention before mark as duplicate. – Antonio Paes May 04 '16 at 14:04
  • it is not clear from your question what is the root cause of the error, and it might still be related to `permute` during HDF5 data creation. Please edit your question to clarify these issues, so it can be considered for re-opening. please add output of `h5ls` on the hdf5 files you are using. If the net is not too long, add a prototxt describing it. you might also want to add more lines from the output from "Setting UP".. of input layer to the error... – Shai May 05 '16 at 08:07

0 Answers0