Connect fully connected layer to a convolution layer?

When building a CNN network for image reconstruction with fully connected layer followed by convolution layers, I found that the output of fully connected layer is always a vector. If I want fully connected layer to be followed by a 2D convolution layer, is there any intermediate layer to reshape the data transferred in between two layers, either on output of fully connected layer or input of convolution layers?

Thanks.

Answer 1

[Uttiya Ghosh]

Hi Kittel,

From my understanding, you want to introduce a reshape layer between your fully connected layer and 2D convolution layer. There is no reshape layer in MATLAB. As a workaround you can always create a custom layer.

For more information, refer to the following links.

https://www.mathworks.com/matlabcentral/answers/461988-reshape-layer-in-deep-learning#accepted_answer_376083

https://www.mathworks.com/help/deeplearning/ug/list-of-deep-learning-layers.html

[Ben Hinkle]

There is now a reshape layer in MATLAB (since R2025a): https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.reshapelayer.html

The syntax to reshape from fully connected to 2D convolution is

reshapeLayer([h w c], OperationDimension="spatial-channel")