Update DataListLoader to reduce code changes ? #3404
QuanticDisaster
started this conversation in
Ideas
Replies: 1 comment
-
I don't think this is possible, as the |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hello,
I would like to know if the following behavior is intended due to limitations :
While I know
DistributedDataParallel
exists in PyTorch,DataParallel
has the advantage of requiring minimal code changes, only requiring to apply it on the model (or specific layers).However, In PyG, using the DataParallel of PyG library necesitates to change
DataLoader
toDataListLoader
in code (up to this point it is still good) but any try to access attributes of a batch results in an error if code is not updated :'list' object has no attribute' pos'
This forces to update all access to attributes (labels for metrics computing), data transforms and so on
In order to reduce the code changes needed, would it be possible to improve
DataListLoader
in PyG such that all methods implemented by (acces to attributes) and on (data transformations for instance) DataLoader objects return the same result without need to change ?An example of the problem :
Beta Was this translation helpful? Give feedback.
All reactions