Checkpoint not working with batch normalization

I am using the Deep Learning Toolbox to build and train a convolutional network. In my network I have batch normalization layers after my convolutional layers. I attempted to save my progress using the CheckpointPath option, but when I loaded my network from a checkpoint via load('checkpoint_name.mat'), using predict(net, XValidation) gave the following error:

"Error using DAGNetwork/calculatePredict>predictBatch (line 157)

Input network must not have batch normalization layers with empty

Mean or Variance properties. To calculate the batch normalization

statistics, use the trainNetwork function.

Error in DAGNetwork/calculatePredict (line 17)

Y = predictBatch( ...

Error in DAGNetwork/predict (line 131)

Y = this.calculatePredict( ...

Error in SeriesNetwork/predict (line 263)

Y = this.UnderlyingDAGNetwork.predict(X, varargin{:});

Error in TestNetworks (line 7)

YPredicted = predict(net, XValidation);"

Does someone know what causes this and/or how to fix it?

Thanks in advance.

Answer 1

[Mahesh Taparia]

Hi

The possible workaround of this problem is to save the weights of the network or the complete workspace after completion of training using save function. While making the inference, load that back to the workspace. Hope it will help!

[Gelin Guo]

Hi,

This is still a standing issue after 2yrs passed in R2022a. Do you have any updates on this issue? Due to this bug, the checkpoint functionality in the trainedNetwork is basically rendered useless when BatchNormalization layer is in use.

[Sam Leeney]

For anyone else stuck, there is a fix here: https://uk.mathworks.com/matlabcentral/answers/423588-how-to-classify-with-dag-network-from-checkpoint