-
Notifications
You must be signed in to change notification settings - Fork 1.8k
Conversation
@penghouwen @Z7zuqer, Hi, can you help to review this pull request? thanks:) |
Hi, @penghouwen @Z7zuqer ~~ |
Hi, Thanks for your interest in our project! Two bugs you mention in this PR are indeed in our current codebase, I think it's ok to merge this PR. Best, |
Hi, @QuanluZhang ~ |
sorry, only the account that joined microsoft org can be assigned as reviewer i guess. I think @Z7zuqer will help review :) |
@Z7zuqer if you think this pr is ok, please approve it |
Hi, I think it's ok to merge this PR, but I am not the reviewer, could I ask @penghouwen to help me approve this? Best, |
you can approve without being assigned as an reviewer :) |
Approved. @AliCloud-PAI Thanks for your interest and kind helps on Cream! |
Hi, @QuanluZhang ~~ |
Two bugs may be fixed for the NAS (Cream of the Crop):
in "nni/nni/algorithms/nas/pytorch/cream/trainer.py", line 168:
I think the prioritized board should be updated with the newest student network, not the teacher. Otherwise, the prioritized board will be filled with the same teacher?
in "nni/examples/nas/cream/lib/models/structures/supernet.py", line 126:
In the class method rand_parameters(self, ...), when yielding parameters of the block in mutable, the passed architecture is the dictionary with the type Dict[str, np.array], of which the key is mutable.key, and the value is a one-hot vector. In the original function, the traversed blocks is already a block in the mutable, which can not be indexed by the arch. Therefore, the class method is modified by using (block, arch) to find the sampled block in the mutable.