scan/scan_op: Convert known_grads to OrderedDict
This was probably causing a different order of operation
during gradient computation in scan for each run. With this
fix I'm able to finally reproduce results on my RNN system.
正在显示
请
注册
或者
登录
后发表评论