2

I am currently working with a very simple CNN and I am trying to figure out a way to make it into a Full CNN. By Full I mean not having any Dense layer and instead implement a deconvnet. The deconvnet part will be used as a localization part of the CNN. What I am wondering is if there is a way to implement a deconvnet part in nolearn? In detail, a Deconv2dlayer and Unpooling? Has anyone else looked more into it or even done it?

Here is a reference for more details:

https://datascience.stackexchange.com/questions/8999/deconvolutional-network-in-semantic-segmentation

Community
  • 1
  • 1
Saudin Botonjic
  • 21
  • 2
  • I try to do exactly the same thing Also based on the same paper and with nolearn. I just started yesterday playing with these things, but as soon as i know the unpooling and deconvolution can be implemented with the InverseLayer which is very easy to use. In general the code looks like in the following question: https://stackoverflow.com/questions/37147516/lasagne-autoencoder-how-do-i-just-use-the-decoder-part – Kevin Meier Aug 30 '16 at 10:09

0 Answers0