Vikas89 commented on issue #14340: [bug] Bug in Gradient flow with backward(retain_graph=True) and split() URL: https://github.com/apache/incubator-mxnet/issues/14340#issuecomment-484304929 Ok, the root cause for this is we have a caching in nnvm::legacy_op_util, on first backward pass the data will be inititalized, on next backward pass, split backward gets junk data. If I defeat caching here: https://github.com/apache/incubator-mxnet/blame/master/src/nnvm/legacy_op_util.cc#L146 by making bwd_init_ = false , I get expected output. Not sure what was the reason to introduce caching and if what is the consequences if we just remove it. @tqchen : by any chance do you remember and if it is safe to remove caching ?
---------------------------------------------------------------- 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
