Validation loss of a model is not stable compared to training loss

[Mohammad Yahya]

Below I have a model trained in Keras and the loss of both the training dataset (blue) and validation dataset (orange) are shown. From my understanding, the ideal case is that both validation and training loss should converge and stabilize in order to tell that the model does not underfit or overfit. But I am not sure about the model below. What you can tell from its loss, please?

In addition, this is the accuracy of the model:

[geantbrun]

I don't know what you call here accuracy here but in theory, it should decrease with the epochs. Try to play with learning rate maybe?

[geantbrun]

*increase* sorry

[Lance Norskog]

The accuracy graph looks like 'validation loss' and 'validation accuracy'. Are you sure you have the right numbers in these plots?

Also- I see this kind of wildly changing accuracy when I use a simple Dropout layer in a CNN. SpacialDropout is "smart" about CNNs and might work better for you.

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/keras-users/6b21b38e-7e2b-4b6b-a524-7f3474c4958cn%40googlegroups.com.

--

Lance Norskog
lance....@gmail.com
Redwood City, CA