You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Using ignite.engine.create_supervised_trainer I'm unable to clip my gradients as this falls between the steps:
loss.backward()
# gradient clipoptimizer.step()
There are no event's fired between these two pieces of code. Gradient clipping and other manipulation techniques are quite common. It'd be nice if Ignite supported these. Is there a recommended way of doing this? Or is it just a case of constructing my own Engine with a process_fn similar to that constructed by create_supervised_trainer? Are there any plans to create events that allow one to do this sort of thing with create_supervised_trainer?
Cheers for the awesome lib!
The text was updated successfully, but these errors were encountered:
Concering your problem I would advise (see Concepts) to use directly Engine with custom process_fn without complicating the system with custom events.
Methods like create_supervised_trainer are just helper methods for a basic usage.
Using
ignite.engine.create_supervised_trainer
I'm unable to clip my gradients as this falls between the steps:There are no event's fired between these two pieces of code. Gradient clipping and other manipulation techniques are quite common. It'd be nice if Ignite supported these. Is there a recommended way of doing this? Or is it just a case of constructing my own
Engine
with a process_fn similar to that constructed bycreate_supervised_trainer
? Are there any plans to create events that allow one to do this sort of thing withcreate_supervised_trainer
?Cheers for the awesome lib!
The text was updated successfully, but these errors were encountered: