Custom Mask-rcnn

Hello, is it possible to make a smaller mask rcnn? that is, to build it on a smaller res-net 50? when training YOLO, i can choose the feature extraction layer. Can I do it also for Mask rcnn? I don't really understand the literature, and the training of the net is taking so long that I was not able to complete a single training session, let alone try to tune the right hyperparameters for my application. I have a set of brigthfield images, [2048x2048], which I downsize to 300x300x3 for the training.

Thank you to whoever will respond.

Answer 1

[Sachin]

Hi

I understand that you want to create a smaller mask RCNN. The following workaround might be of good help to you.

You can use ‘removelayers’ to remove a particular layers and “addlayers” to add a layers.

Here is an example which you can refer:

net = resnet50; % resnet object model
lgraph = layerGraph(net); % architecture of a deep learning network
lgraph = removeLayers(lgraph, {'fc1000', 'prob', 'ClassificationLayer_predictions'}); % to remove a layer
feature_layer = 'activation_40_relu'; % feature extraction layer
lgraph = connectLayers(lgraph, 'res5c_relu', feature_layer); % connect all the layers

Refer the following MATLAB documentation page for more information about ‘removelayers’

https://in.mathworks.com/help/deeplearning/ref/nnet.cnn.layergraph.removelayers.html#d124e164785

Refer the following MATLAB documentation page for more information about ‘addlayers’

https://in.mathworks.com/help/deeplearning/ref/nnet.cnn.layergraph.addlayers.html#d124e10823

For more information also refer this MATLAB answer

https://in.mathworks.com/matlabcentral/answers/587168-convert-image-classification-network-into-regression-network-using-resnet18

Thanks