Convolution Neural Network Checkpoints with Batch Normalization

The main problem that I'm addressing is that checkpoints from trainNetwork() are broken if they have Batch Normalization layers. I think this is a known issue.

I was thinking of running trainNetwork() again on each checkpoint for a single epoch on a small subset of my training images with realmin learning rate just to finish the training without affecting the network too much.

The issue i'm having is setting the average image property in the imageInputLayer to be static. I already have the zerocenter mean value for my entire training dataset and that will be overwritten with the stats from the smaller dataset which is read-only. Do i really have to use the same training set of over 1 mil images at a realmin learning rate to make my checkpoints usable?

Is there another approach i should try?

[Sam Leeney]

Infuriating that a product people actually pay for has left a fundamental bug like this unfixed for so long. The checkpoint function is rendered useless. Using MATLAB is becoming more infuriating every day.