For IAM data training, after how many epoch, does the network yield reasonable error rate? about returnn HOT 10 CLOSED

Hi,

first you should determine if it's a problem of overfitting or underfitting. You can use the same data for training and validation to see if the error rate on the training set goes towards 0. If this happens, then you have overfitting, which probably means that your trainset is too small. If you don't even fit the training data well, then something is broken. In this case, maybe try to use the normal split of the dataset and the provided config file and see if you can reach a good result.

from returnn.

@pvoigtlaender
Thanks! I will try that.
What about LEARNING rates?
Any justification for using 0.0005, 0.0003, 0.0001 ?
I see there is code that calculates error rates from epoch!
Thanks again for help out!

from returnn.

I think you are talking about learning rates, not error rates, right? We determined them empirically, so basically trial and error. For a different setup it might be useful to adapt it.

from returnn.

Sure, how long does it take your research group to train the entire IAM dataset (with config_real) ?
And roughly how many epochs?
Is gtx 1080 ti + i7 machine up to the task?

from returnn.

GTX 1080 TI + i7 is definitely enough! I don't remember the exact values anymore, but I think we trained for ~80 epochs and 1 epoch might take around half an hour (very roughly).

from returnn.

Thanks!
Great Help!

from returnn.

If for some reason, I had to interrupt the training, I could always continue training from where I stopped by setting the "start_epoch" parameter in the configure file to "last saved epoch +1".
This should have no effect on the whole training result, right!
I hope not!
Thanks!

from returnn.

I think you can just rerun without any changes to the config and RETURNN will load the most recent model.

from returnn.

Got it!

from returnn.

Not sure that this is about a bug in Returnn, so closing.

from returnn.