Custom flattenLayer compatible with imageInputLayer

I have been reading through some papers as part of research for my BSc major project, and came across this paper https://www.ijeat.org/wp-content/uploads/papers/v8i6/F8602088619.pdf outlining a hybrid LSTM-CNN architecture. While not the main subject of my project, I am curious about implementing the basic diagram that they have listed.

The architecture is simple enough: image input, batch normalization, lstm, convolution, max pool, fully connected, output.

Obviously a flatten layer is needed between batch norm and lstm, however the flatten layer provided in matlab is not compatible with image input layers (both 2D and 3D).

Reading the Flatten.m source file, the comments list the basic details of image dimensions, however the FlattenLayer.m (class) and flattenLayer.m (function), only list sequence data dimensions in their comments.

This makes me wonder if it is at all possible to make a custom flatten layer that is compatible with image input.

Answer 1

[Madhav Thakker]

Hi Amelia, I understand that you want to use flatten layer with Image input layer. As of now, faltten layer supports sequence input only. This exhancement is being taken into consideration and the concerned parties may be investigating further.

You can always define your own custom deep learning layers if required. https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html

Hope this helps.

[Shino Asada]

I have a simple CNN network that uses FlattenLayer, which prevents me from porting the network to MATLAB to run. If MATLAB implements ReshapeLayer, I believe this will reduce a lot of trouble.

Answer 2

[Dan]

same ask. it will be very beneficial from compatability with other frameworks