0
$\begingroup$

I'm trying to make a dark image brighter using CNN-UNet architecture.

When I train the network, I get the following results:

enter image description here

When I cut the features in half for pruning, and do full train again, I get the following result:

enter image description here

There are those bubble artifacts near the light.

Why are they happening?

  1. How can I locate the problematic neurons that cause it?

  2. Can I train the network (change the loss or anything) to not generate those made up bubbles?

$\endgroup$
1
  • $\begingroup$ This is an old question, but what do you mean by "pruning"? Moreover, what do you mean by "U-net CNN"? Do you mean just the regular U-net? $\endgroup$ – nbro Feb 3 at 10:11
0
$\begingroup$

Did you retrain the network after pruning the weights?

If not, 1) the bubbles are probably appearing because of that, and 2) you definitely should retrain your network.

Did you prune output features only or weights inside the network also?

The standard procedure is to retrain the network after pruning, and the best results can be achieved with iterative pruning: you prune small part of weights (the loss of accyracy is inevitable here), then retrain to get the accuracy back, and repeat until you won't be able to get the accuracy back to high level.

Here Pruning Convolutional Neural Networks for Resource Efficient Inference is an article about (among others) iterative pruning.

$\endgroup$
1
  • $\begingroup$ I do full training and get those results. They occur becuase I changed number of channels from 16 to 8 in the network. $\endgroup$ – BestR May 14 '19 at 14:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.