How to train with Async SGD?

[zuowang]

I am confused on how to train with Async SGD.
I found there is a command line argument --async_count, but it's not used at all.
Could you give me an example?

[backyes]

check setting(...) in model config file.

python/paddle/trainer_config_helper/optimizers.py
338 def settings(batch_size,
339 learning_rate=1e-3,
340 learning_method=None,
341 regularization=None,
342 is_async=False,
343 model_average=None,
344 gradient_clipping_threshold=None
345 ):

368 :param is_async: Is Async-SGD or not. Default value is False.
369 :type is_async: bool

So, set is_asyn=True enable async-sgd. For most model, is_asyn=True is preferred

[zuowang]

Thanks!