Invalid 'ValidationData' argument prevents other trainingOptions from loading properly

2 次查看（过去 30 天）
Gabalodon
Gabalodon 2023-4-21
Hello!
I created and trained a YOLOv4 object detector in which 'BatchNormalizationStatistics' is set to 'moving', 'ResetNormalizationLayer' is set to false, and 'ValidationData' is set to a datastore I created using about 10% of my training data. The training goes well and I save the resulting 'detector' and 'info' from the trainYOLOv4ObjectDetector function, as well as the 'options' (result from 'trainingOptions') used to train the network. If I then load these variables into an empty Matlab environment, everything goes fine, but if I relocate the directory containing the original datastores used to train the network before loading them, I get the following warning:
"Warning: While loading an object of class 'nnet.cnn.TrainingOptionsRMSProp': The value of 'ValidationData' is invalid. Unable to read file: 'Path\to\pngs\used\in\original\datastore'."
This causes ValidationData to be set to [], which would be fine and I expect that behavior. Unfortunately, all of my other trainingOptions ('BatchNormalizationStatistics', 'ResetNormalizationLayer', etc.) get reset to defaults, which prevents me from being able to build on the pre-existing network. This seems like it shouldn't be the case, especially since it is a 'warning' not an 'error'.
Would it be possible for Matlab to resolve this so that ValidationData gets reset to [] but everything else stays the same?
Thank you!

请先登录，再回答此问题。

回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Deep Learning Toolbox 的更多信息

标签

产品


版本

R2022b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by