mk-61 opened a new pull request #20718: URL: https://github.com/apache/incubator-mxnet/pull/20718
Automatic Layout Management improves performance when used together with AMP by converting parts of the computational graph to NHWC layouts. ## Description ## Target graph parts (containing convolution / deconvolution ops) are automatically converted to NHWC layout by surrounding with transposes. This functionality was originally implemented by Dawid Tracz <[email protected]> in the Nvidia container. I later changed the algorithm to a single pass of DFSVisit. ## Checklist ## ### Essentials ### - [X] Changes are complete (i.e. I finished coding on this PR) ### Changes ### - [X] A new layout_optimization flag is added to amp.init() call ## Comments ## This other PR: https://github.com/apache/incubator-mxnet/pull/20635 is required in order to get the expected performance improvements. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
