[Relay][FrontEnd][MxNet] incorrect data_min_idx and data_max_idx in _qnn_conv

[GuoliangLiCN]

In the lines 1409-1413 of file python/tvm/relay/frontend/mxnet.py, are there typos when has_sum is False?

        data_min_idx = -1
        data_max_idx = -2
        if has_sum:
            data_min_idx = -4
            data_max_idx = -3

Should data_min_idx be -2 and data_max_idx be -1 according to lines 467-474 of https://github.com/apache/incubator-mxnet/blob/master/src/operator/subgraph/mkldnn/mkldnn_conv.cc

if (param.full_conv_param.mkldnn_param.quantized) {
input_names.emplace_back("data_min");
input_names.emplace_back("data_max");
if (param.full_conv_param.mkldnn_param.with_sum) {
input_names.emplace_back("sum_min");
input_names.emplace_back("sum_max");
}
}

[tqchen]

cc @kevinthesun @zhiics can you help check on this ?

[zhiics]

yeah, it looks flipped to me as well. @shoubhik can you confirm? Thanks.

[shoubhik]

taking a look at it.

[anijain2305]

Thanks for finding the bug @GuoliangLiCN
Yeah, it is flipped. The bug hid because the scale and zero point remained same even if the data_min and data_max were flipped.

[shoubhik]

yes, it does needs a flipping. the test case passes because the scale value remains the same. I'll send a PR a little later.

[tqchen]

What is the state of this issue? @anijain2305 @shoubhik

[shoubhik]

sorry got swamped. I will definitely send a fix on Monday.

[tqchen]

#5136