Problem with pre trained Resnet 50

Hi, i was trying to use pretrained resnet50 to train my own dataset. But i got the following error "The output of layer 12 is incompatible with the input expected by layer 13" So far i know to use a pretrained network, only last three layers need to change. i have change those but still getting the same error.

Can anyone help me on that?

[Chuanzhen Hu]

I also met this problem, do you know how to solve it now?

[Noha Ghatwary]

I have the same problem ,, did you manage to find a solution ???

[SAQIB SHEIKH]

Every Pratined Model accpect fixed size of all images thats's why problem occur you should be read about the Resnet50 than apply on your data.

image input size of 224-by-224. for This Model

Answer 1

[Chuanzhen Hu]

https://www.mathworks.com/help/nnet/ref/resnet50.html check this web, it says, 'you can use classify to classify new images using the ResNet-50 model. To retrain the network on a new classification task, follow the steps of Transfer Learning Using GoogLeNet', ResNet-50 is a DAG net, it's different with series net.

[Soan Duong]

FYI: https://au.mathworks.com/help/nnet/examples/transfer-learning-using-googlenet.html

Answer 2

[Divas Bahguna]

This error normally occurs when you try to re-build ResNet50 as a sequential layered network after replacing classification layers with your custom layers. As highlighed by Chuanzhen Hu, ResNet50 is a DAG in Matlab, therefore requiring the reconstructed network to be a layer graph. You usually have to consider connectivity relationships between layers and have to re-construct the network by factorign these relationships in layer graph.

New MATLAB versions come with library APIs to take care of these reconstuctions. Eg.,

https://in.mathworks.com/help/deeplearning/ref/replacelayer.html

However, if you are using older versions of MATLAB you'll have to write these implementations yourselves. Hope this solves the issue.