shoubhik commented on issue #4764: [CI] ci-gpu update blockers URL: https://github.com/apache/incubator-tvm/issues/4764#issuecomment-581189877 > given that the mkl part poses accuracy problem, i feel it might be a bad idea to rely on it for testing QNN(see also comment about intel dependency). would be great if we can explore generic alternatives that can test QNN. For the parser part, I think we can start by directly checking alpha equivalence of the graph as well as potentially the comparison to a simulated FP32 version. + @anijain2305 @tqchen most of the code for mxnet qnn has been tested for MKLDNN. Some of the formulas used for quantization, dequantization and convolutions used my the reference Mxnet implementations(fake quantization) is not fully and thorougly tested. Also some of the optimizations available in MKLDNN are not implemented in the stock implementaion. So, my suggestion to unblock this PR would be to use the simple feed forward network test case, where we test only the graph generated by QNN. Once we have the fix from MKLDNN as well we can add proper test cases with MKLDNN at that time. @tqchen and @icemelon9 what do you think?
---------------------------------------------------------------- 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
