Proper training stopping without max_steps configuration

[jonmay]

What data and parameters were used to build the pretrained model and the model used to generate PRONALSYL 2007 numbers in row "LSTM num_layers=2, size=512" in the table? The table claims 23.3 WER, when I run the pretrained model on the test set of PRONALSYL 2007 I get 22.8, and when I run my own model trained according to the README in the PRONALSYL download (--train cmudict.dic.train --test cmudict.dic.test --num_layers 2 --size 512 --max_steps 0) I get 32.11. Is the data the same for all these? Are there any tricks I am missing? Are random restarts needed?

[nshmyrev]

You need more steps, something like --max_steps 70000, we are currently working on proper stopping.

[jonmay]

ah, thanks, i'll give it a try