New LSTM error! about paddle HOT 4 CLOSED

There are several things you can try:

1. decrease learning_rate
2. Try a different optimization algorithm
   These can be changed in settings(). See http://www.paddlepaddle.org/doc/ui/api/trainer_config_helpers/optimizers.html#settings

from paddle.

hi @emailweixu , so your meaning is that larger learning_rate or some optimization algorithms will lead to floating point exception? And why is this? Are there some parameter updates too large or zero ? THX~

from paddle.

Hi @NIULQfromNJU . I had similar Issue #46 with LSTMs (internal state grow over time), switched to GRUs and have no issues with them so far. Judging by "An Empirical Exploration of Recurrent Network Architectures" paper they are comparable in terms of accuracy. GRU are stable, because the state cannot grow higher than the maximal input

from paddle.

@F0REacH thanks man~

from paddle.