How to build a single-object localization CNN based-on an object classification network?

Hi all,

I'm new to convolutional neural network. I'm trying to use Alexnet to build a single object localization network, which I think is a simple problem in this area. What I did is to replace the last three layers of Alexnet, which are fullyConnnectedLayer(1000), softmaxLayer, and classificationLayer, with two layers, which are fullyConnectedLayer(4) and regressionLayer. Did I do it right? When I tried to train this network, the result seems wrong. Specifically, the Mini-batch Loss and the Mini-batch RMSE were NaN all the time. What would be the reason for this?

Thank you in advance!

Shuwei.

Answer 1

[Ankitha Kollegal Arjun]

If the regression response values are very big and/or you have many regressors, your loss will become large, and it eventually leads to NaNs. This is the reason your Mini-batch Loss and the Mini-batch RMSE are NaN.

There are two workarounds which might fix the issue. Try one of the following:

1. Lower the initial learning rate. See more details on the initial learning rate here:

Initial Learn Rate

2. Normalize the responses so that the maximum value is 1. You can use the function "normc" for this:

Normalize

Hope this helps!

[Shuwei Qiu]

Thanks for your answer! I fixed this NaN issue using your first solution. But I have another issue which is the predicted bounding box from my network is always the same no matter what is the input image. I was wondering do you have any idea to fix this? Thank you very much in advance!