[Need Help] Cannot use DL checkpoint model to predict?
1 Ansicht (letzte 30 Tage)
Ältere Kommentare anzeigen
I tried to use models saved in path specified by 'CheckpointPath' parameter during training by Deep Learning Toolbox, I got following error:
Error using DAGNetwork/calculatePredict>predictSingle (line 112)
Input parameter has the wrong class.
Error in DAGNetwork/calculatePredict (line 13)
Y = predictSingle( ...
Error in DAGNetwork/predict (line 125)
Y = this.calculatePredict( ...
The detector after training can be used without problem.
Did I miss something? Are the checkpoint models of same class as final detector itself?
Thanks
2 Kommentare
Guillaume PERRIN
am 29 Mär. 2020
Bearbeitet: Guillaume PERRIN
am 29 Mär. 2020
Hi, indeed checkpoints that are associated to some DAGNetwork architectures cannot be used to make predictions. As an example, those who contain batchNorm layers. See https://fr.mathworks.com/matlabcentral/answers/423588-how-to-classify-with-dag-network-from-checkpoint or https://fr.mathworks.com/matlabcentral/answers/451383-issue-with-batch-normalization-layer-of-saved-cnn#answer_366855.
For a workaround, you can still retrain from your checkpoint during 1 epoch, with a minimalist training set, at a very low learning rate.
Best,
Guillaume
Antworten (0)
Siehe auch
Kategorien
Mehr zu Deep Learning Toolbox finden Sie in Help Center und File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!