zhreshold commented on issue #16708: Training an FPN model using grad_req="add" causes rapid divergence, while manually implemented gradient accumulation works fine URL: https://github.com/apache/incubator-mxnet/issues/16708#issuecomment-558876214 After digging a while, I found several confusing facts about this bug. 1. @nickguletskii is correct, it's not about `ElementWiseSum` 2. Duplicating same node with (1, 2, 3, 4, 8, 9, 10...) times, the loss and gradients are always GOOD, however, with (5, 6, 7), the gradients will diverge at the first iteration
---------------------------------------------------------------- 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: us...@infra.apache.org With regards, Apache Git Services