Open QVQZZZ opened 2 months ago
Weight slicing for OD layers has the input and output dimensions reversed.
In the weight parameters of Linear and Conv layers, the first dimension should be out_dim, and the second dimension should be in_dim.
out_dim
in_dim
Reversed the order of input and output dimensions in the weight slicing for OD layers.
#contributions
Thanks for opening the PR @QVQZZZ , let me loop in @stevelaskaridis and @SamuelHorvath who are the masterminds behind FjORD and added the baseline: what do you think about this 2-lines change?
Issue
Description
Weight slicing for OD layers has the input and output dimensions reversed.
In the weight parameters of Linear and Conv layers, the first dimension should be
out_dim
, and the second dimension should bein_dim
.Related issues/PRs
Proposal
Explanation
Reversed the order of input and output dimensions in the weight slicing for OD layers.
Checklist
#contributions
)Any other comments?