sxjscience commented on issue #17703: Need to use safe accumulation for calculating the gradient of Embedding + Take URL: https://github.com/apache/incubator-mxnet/issues/17703#issuecomment-591825658 To clarify, it affects both nd and the new numpy (I'm testing with the numpy interface because I'd like to test the `take` operator). This mimics the setting that appears in large-batch training of transformers, in which there will be way more tokens than the size of the vocabulary.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
