Marcel
Hello everybody,
a couple of month ago I started to work with Caffe and image segmentation with a changed GoogleNet. So far most of what I have done was training networks, tweak parameters in the solver.prototxt and change some layers. I plotted my results with the plot_log.py script from the Caffe toolbox to check the networks performance.
But after reading the Stanford course cs231n I want to look deeper into the network. I changed my accuracy layer, so I would get training accuracy score as well. But I have trouble realizing the “Ratio of weights:updates”, “Activation/Gradient distributions” per layer and “visualizing the first layer”.
Right now I try to adapt the Caffe tutorial „Classification: Instant Recognition with Caffe“ to show me the first-layer Visualization. This tutorial also shows a plot for the “output values” and the “histogram of positive values”. Sadly it lacks any kind of documentation. What do these Graphics show and are these maybe already the activation/gradient histograms I am looking for?
If anybody already solved one of my 3 problems I would be most thankful for any help :-)