Earlier this month, the Neural Network Toolbox team submitted a new Deep Learning Network Analyzer tool to the File Exchange. (Note: it requires the R2018a release.) This very useful tool helps you spot problems if you are building a network from scratch, or if you are modifying the structure of an existing network. It can identify missing or disconnected layers; mismatched layer input sizes; incorrect number of layer inputs; or invalid graph structures.
To try it out, I loaded in GoogLeNet, and then I started messing with it. I disconnected a couple of layers, and I chopped off the first layer. Then I ran
You can see how the analyzer has flagged the structural errors.
The analyzer also shows you useful details about the individual layers. Here is part of the display for the unmessed-with GoogLeNet:
So, give the new analyzer a try. Just remember that you need the latest release, R2018a, to run it.
To leave a comment, please click here to sign in to your MathWorks Account or create a new one.