Problem with training FCN for 4 channels

[Solitarysea]

Hi,

I am trying to train FCN on 4 channel data. I am using the model designed for Nyud V2. The trouble is that after 100K iterations loss does not decrease. Infact when i plot it is straight line. It is also quite large 3e7. I have a suspicion that it might not be due to not supplying a singleton dimension in the groudtruth. More specifically, as far as i understood from the comments that i have to provide a singleton dimension in addition to 2D image (i.e. 1xHxW). But if i provide the FCN with this data then i run into problem of Nd convolution. I get the following error in that case:

F0713 14:12:04.008546 32691 blob.hpp:140] Check failed: num_axes() <= 4 (5 vs. 4) Cannot use legacy accessors on Blobs with > 4 axes.

Would anybody please verify that i must have to provide the singleton dimension and/or how to fix the static loss problem?

Best,

[dksa...@gmail.com]

Could you share your prototxt file?