Network slower with less classes in last layer

[Ludwig]

Hi,

I am facing the issue that training a network became several times slower, after I switched to a network with fewer classes. Did anyone experience that before or has an idea what could be the problem?

Both datasets exist as lmdb files with the images reduced to the same size.

Batch_size, crop_size and the other parameters are the same except for the num_output of the last (fully connected) layer.

Thanks

Ludwig

[Ludwig Schmidt-Hackenberg]

p.s. I know, it is a strange problem and I assume, that I have overlooked something trivial, like not resizing the images for the lmdb file. I am thankful for any hints.

--
You received this message because you are subscribed to a topic in the Google Groups "Caffe Users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/caffe-users/SEqyniq9s6s/unsubscribe.
To unsubscribe from this group and all its topics, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/17f229e8-aac3-4427-83d3-174a98a31782%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

EyeEm

Ludwig Schmidt-Hackenberg

Computer Vision Engineer

@shackenberg

EyeEm Mobile GmbH
Kohlfurter Strasse 4I/43, I0999 Berlin, Germany
Geschäftsführer: Florian Meissner & Lorenz Aschoff
Handelsregister Amtsgericht Charlottenburg HRBI3253I