Are there any convolutional networks in the DL Toolbox whose output layers are implemented as fullyConnectedLayers?

Within the library of neural networks offered through the Deep Learning Toolbox, are there any convolutional neural netwoks that contain fully connected layers? If, as an exercise, I wished to convert the fully connected layers to convolutional layers (so as to convert the network to an FCN), is the procedure for doing so straightforward?

[Matt J]

cf. also

https://www.mathworks.com/matlabcentral/answers/880863-convnets-and-sliding-windows?s_tid=srchtitle

Answer 1

[David Willingham]

Hi Matt,

In this doc page, it shows an example of a network that has both convolutional layers and full connected layers:

https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.fullyconnectedlayer.html

Can you clarify a little more on why you want to convert fully connected layers to convolutional layers?

David

[Matt J]

Thanks David,

Can you clarify a little more on why you want to convert fully connected layers to convolutional layers?

Essentially, I would like to know how to adapt a network so that it becomes fully convolutional. A fully convolutional network has a property that I would like to exploit. Namely, it can be trained on images of one size (NxN), but then at inference time, it can be given an input image A of a larger size (MxM). In the latter case, the output will be what you would get if you were to extract all NxN sub-images of A, put them through the network one at a time, and concatenate the results. However, doing that literally is much less efficient than just passing A through the network once, in its entirety.

[Matt J]

Also, regarding the doc page you referenced, I know that it is possible to construct a CNN wih fully connected layers, but my question is about the stock networks that I see when I open deepnetworkDesigner. In the deepnetworkDesigner library, there don't seem to be any CNNs with fully convolutional layers. I'm wondering if that is coincidence.

[David Willingham]

Is this what you're looking for?

https://www.mathworks.com/matlabcentral/fileexchange/65851-semantic-segmentation-using-fcn-alexnet?s_tid=srchtitle_fcnlayers_3