option for non-strict preloading ckpt #1785
Conversation
| @@ -62,6 +62,7 @@ def _flatten_module(module): | |||
| return [module] | |||
|
|
|||
|
|
|||
| @alf.configurable(whitelist=['checkpoint_preload_strict']) | |||
There was a problem hiding this comment.
I feel it is better to pass this argument from the subclass algorithms and config for the subclass algorithm so we don't make global change for the behavior of Algorithm.
There was a problem hiding this comment.
I feel it is better to pass this argument from the subclass algorithms and config for the subclass algorithm so we don't make global change for the behavior of Algorithm.
I think the question is down to which subalgorithm? If all the way down to each leaf algorithm, there will be lots of argument change lines across the entire repo.
Or I can add this option to TrainerConfig so that it's easier for the change.
There was a problem hiding this comment.
I feel it is better to pass this argument from the subclass algorithms and config for the subclass algorithm so we don't make global change for the behavior of Algorithm.
I think the question is down to which subalgorithm? If all the way down to each leaf algorithm, there will be lots of argument change lines across the entire repo.
Just change the specific algorithm that needs this behavior at this moment.
While currently we allow the user to configure whether a ckpt is strictly loaded or not in
policy_trainer.pyfor loading an algorithm's entire ckpt, there is no such an option for in-alg ckpt preloading. This PR adds a flag toAlgorithmto allow partial loading. This is useful if we want to ignore loading some model weights that can be downloaded as pretrained ones.