why is model in eval mode in train about mean-shifted-anomaly-detection HOT 4 CLOSED

I am not sure but I think one of the reasons might be to remove the effect of Batch-normalization and Dropout. If it was in train mode, the centers would have been heavily affected by the changes introduced by BN.

This is a logical explanation. The authors can confirm the same.

from mean-shifted-anomaly-detection.

But if model.eval() is on, the model’s parameters will not be updated.

from mean-shifted-anomaly-detection.

They do get updated. The nature of some of the layers changes based on eval vs train.
You can train an MNIST with the model in eval mode. If weights were not updating, the loss would not have decreased.

from mean-shifted-anomaly-detection.

Hi,

Indeed the reason for us training the model on eval mode is to remove the effect of batch norm and dropout layers.

from mean-shifted-anomaly-detection.