nirvanjhurreepriv commented on PR #2524: URL: https://github.com/apache/systemds/pull/2524#issuecomment-5117319765
Hey David @ywcb00 , On the reconstruction bound: it was a worst-case analytical estimate: for TopK, zeroing the smallest weights of B changes each result entry by less than the sum of those weights; for quantization, each weight shifts by at most one level step - with a safety factor on top. But I removed it. In its place I added unit tests that check compress/decompress accuracy directly on known inputs (TopK preserves the largest values exactly; quantization stays within one step), which is exact rather than bounded and also covers the receive-path code. The reference script now differ from the federated one - the two-worker reference does rbind(read($in_X1), read($in_X2)) in DML, following the FederatedReaderTest pattern. That also let me drop the manual matrix concatenation in Java entirely; the halves are recombined in DML now. Also removed the HTML symbol - was a bug. Thanks for your time David. Best, Nirvan -- 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]
