lrn implementation is different from definition from onnx website.

[scxiao]

in the file test/lrn_test.cpp, the code is as follows:

          auto radius = (lrn.GetN() - 1) / 2;
            par_ford(n_batch, height, width)([&](int b, int h, int w) {
                double scale = 0;
                ford(channels)([&](int c) {
                    auto start = (c - radius) < 0 ? 0 : (c - radius);
                    auto end   = (c + radius) > channels ? channels : (c + radius);

                    for(auto k = start; k < end; k++)
                    {
                        scale += std::pow(input(b, k, h, w), 2);
                    }

But the formula for this operator is:
max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2))}

which means the end should be as:

radius_upper = lrn.GetN() / 2 + 1;
auto end   = (c + radius_upper) > channels ? channels : (c + radius_upper);

[scxiao]

the following test case can trigger this bug:

data = 1.24726, 0.282953, 0.692061, 1.58436, 1.32044, 0.691029, 0.671802, 0.107024, 0.072982, 0.290336, 0.101854, 1.02826, 0.538117, 0.086037, 0.313857, 0.235939, 1.0564, 0.145062, 0.221358, 0.603927, 1.80503, 0.462163, 0.611235, 0.524762, 1.60482, 1.02829, 0.032356, 0.132356, 1.08368, 1.0389, 0.109562, 0.267397, 0.62775, 1.40053, 0.246009, 0.110555, 0.869892, 0.850173, 0.273867, 0.213759, 0.741422, 1.00399, 1.20645, 1.3793, 0.360373, 0.309728, 0.88583, 0.549744, 1.14111, 0.150417, 0.633155, 0.273577, 0.36119, 0.83478, 0.0199271, 0.302868, 0.047034, 0.213876, 0.448483, 0.256486, 0.475469, 0.25423, 0.574961, 1.22058, 0.125457, 0.284095, 1.15261, 0.327219, 0.641497, 0.0367307, 0.515534, 0.0912023, 0.80561, 0.873093, 0.284497, 0.484195, 0.615719, 0.409293, 0.220292, 0.039714, 0.823812, 0.63697, 0.329268, 1.08617, 1.05227, 1.34043, 0.833491, 0.127223, 0.757038, 0.745595, 0.285085, 0.864326, 0.147272, 0.690563, 0.251985, 0.499613, 0.00742448, 1.2627, 0.0897355, 0.284246, 1.3315, 0.952969, 0.898339, 0.685455, 0.829504, 1.37423, 0.292468, 0.528524, 1.35959, 1.04681, 1.32041, 0.640647, 0.608971, 1.35051, 0.189503, 0.567415, 0.669782, 0.109609, 0.43421, 0.652088, 0.26617, 0.777378, 0.210875, 0.937869, 0.491128, 0.105808, 0.307684, 1.30754, 0.475375, 0.288118, 0.544403, 0.381297, 0.476812, 0.0225071, 0.449607, 0.478305, 0.407684, 0.147287, 0.280018, 0.772877, 1.05443, 0.310691, 0.117851, 0.449021, 1.68498, 0.667828, 0.645453, 0.789824, 0.930454, 0.326383, 0.0482537, 1.21144, 0.526582, 0.584367, 0.0696162, 0.469141, 0.796606, 0.763601, 0.811298, 0.30958, 0.352143, 1.36427, 0.671331, 0.0619073, 0.86644, 0.597009, 0.707244, 1.09227, 0.840019, 0.22409, 0.651126, 0.225369, 0.605844, 0.460292, 0.731283, 0.481955, 0.568062, 0.487567, 0.322103, 0.0123595, 0.250298, 0.972186, 0.455091, 1.57202, 0.442101, 1.13276, 0.780851, 0.0368853, 0.522943, 1.09099, 0.914149, 0.18883, 0.027776, 0.825858, 0.37001, 0.121298, 0.545731, 0.58229, 1.5295, 0.944986, 0.261048, 0.169265, 0.777554, 0.463334, 0.452638, 1.14327, 0.0172006, 0.521856, 0.197919, 0.0694008, 0.643551, 0.224294, 0.555972, 0.329803, 0.667783, 0.289938, 0.0120352, 0.2681, 1.5974, 0.0298801, 0.675913, 0.244638, 0.327801, 0.340427, 1.08939, 0.0447329, 0.110667, 0.164173, 0.422361, 0.168235, 1.00689, 0.348828, 0.383856, 0.294191, 0.817526, 0.552377, 1.05672, 1.46358, 0.301408, 0.478635, 0.450732, 0.280913, 0.0939605, 0.210552, 0.218505, 1.18504, 0.814802, 0.763388, 0.575127, 1.03684, 0.368446, 0.40714, 0.100376, 0.225783, 0.488953, 0.491258, 0.513061, 0.978147, 1.11922, 0.4316, 0.840625, 0.358365, 0.421645, 0.0371704, 1.36903, 0.133486, 0.370444, 0.0625183, 0.219825, 0.0688723, 0.282162, 1.96027, 1.38294, 0.275836, 0.461318, 0.276436, 0.349109, 0.0820978, 1.4358, 1.45969, 0.0781631, 0.721331, 0.489319, 1.08629, 0.202474, 0.430505, 0.739075, 0.856365, 0.487767, 0.920459, 0.444122, 0.340135, 1.62897, 0.749529, 0.0961309, 0.803877, 0.0690924, 0.412181, 0.282448, 0.261661, 0.923824, 1.17242, 0.0835475, 0.480918, 0.47116, 0.325774, 0.943397, 0.952227, 0.490537, 0.112822, 0.0945401, 0.762012, 0.796747, 0.516654, 0.27213, 0.0667149, 0.0298195, 0.20284, 0.0435746, 0.075876, 0.508821, 0.574865, 0.194111, 0.629959, 0.818352, 0.220815, 0.111478, 1.59556, 0.498286, 0.666967, 0.528333, 0.840619, 0.546748, 0.837101, 1.88008, 0.428726, 1.24151, 0.318846, 0.48366, 1.17342, 0.755535, 0.320591, 0.486369, 0.85846, 0.311772, 0.198239, 0.25787, 0.110805, 0.409068, 0.247237, 0.540325, 1.01664, 0.964758, 0.487506, 0.461231, 0.368531, 1.30314, 0.337974, 0.339159, 0.438605, 0.493877, 0.00266628, 0.658905, 0.240389, 0.0110884, 0.113791, 0.134812, 0.279198, 0.189311, 0.797605, 0.198302, 0.702231, 0.595116, 0.176393, 0.0349982, 0.349192, 0.454879, 1.11054, 0.146303, 0.62234, 1.20068, 0.273845, 1.59479, 0.723009, 0.0273156, 1.17138, 0.696843, 1.04071, 1.16535, 0.116121, 0.40113, 0.157455, 0.249913, 1.14293, 0.20635, 0.538443, 0.606633, 0.806804, 1.03699, 0.602838, 0.423308, 0.789025, 0.542017, 0.251935, 1.25046, 0.25135, 0.575925, 0.0416626, 0.13085, 0.571079, 1.02277, 0.565875, 0.218569, 0.165076, 1.22516, 0.484008, 0.262206, 0.100453, 1.07471, 1.21586, 0.657243, 0.411686, 1.48094, 0.0874806, 0.0919984, 0.0664333, 0.666803, 1.93698, 0.402561, 0.190841, 0.330085, 1.00186, 0.614393, 0.195778, 0.686656, 0.2226, 0.580927, 0.00374245, 0.566078, 0.0553353, 0.279439, 0.819801, 0.0607616, 0.137368, 0.619268, 0.0813866, 0.323437, 0.682022, 0.553392, 0.078053, 0.745709, 0.579971, 0.327436, 0.197345, 0.23964, 1.42898, 0.331519, 1.55649, 0.140926, 0.0357811, 0.365896, 0.692092, 0.310551, 0.128224, 0.355543, 1.70566, 0.679152, 0.560801, 1.61811, 0.177802, 1.42572, 0.381446, 0.194924, 0.501835, 1.22951, 0.703118, 0.932733, 0.623954, 0.798019, 0.350724, 0.545424, 0.727908, 0.642532, 0.300037, 0.60994, 1.87758, 1.07003, 0.39112, 0.0323176, 0.155922, 0.728191, 0.247443, 0.777995, 0.917734, 1.90629, 0.0522691, 0.465662, 0.363616, 0.719846, 0.0550488, 0.270628, 0.0242126, 0.775193, 0.1656, 0.245681, 0.410982, 1.15439, 1.10851, 0.833771, 0.920198, 0.633033, 0.972215, 0.941987, 1.39194, 0.466727, 0.124323, 0.352617, 0.741014, 0.201008, 1.23219, 0.157405, 0.645623, 1.18862, 0.628575, 0.171202, 0.62841, 0.662367, 0.998599, 1.67539, 0.610964, 1.58345, 0.283863, 0.866057, 0.0458577, 0.904818, 0.413927, 0.185005, 0.128863, 0.143453, 0.0776976, 0.150953, 0.854565, 0.171129, 1.07353, 0.27197, 0.313837, 0.762333, 1.80933, 0.835322, 0.446817, 0.115913, 0.0681077, 0.66638, 0.1892, 0.479417, 0.917555, 1.67155, 0.0143778, 0.953088, 0.538491, 1.42206, 0.0315337, 0.137934, 1.25968, 0.515428, 0.138986, 0.250844, 0.436201, 0.00610049, 0.372645, 0.320869, 1.29372, 0.0261641, 0.542963, 0.417103, 0.257276, 0.569656, 0.790726, 0.092669, 0.801159, 1.38, 0.466567, 0.805908, 0.555026, 0.391949, 0.332751, 0.153402, 0.314937, 0.277442, 2.1536, 0.384176, 0.310448, 0.155238, 0.766493, 0.248735, 0.268156, 0.332343, 0.153252, 0.657703, 0.12628, 1.09627, 0.295085, 0.667746, 0.168364, 0.994112, 0.417205, 0.0781272, 1.17421, 0.081421, 0.268096, 1.23196, 0.921505, 0.42788, 0.633244, 0.093259, 0.286208, 0.158281, 0.233078, 0.909285, 1.06557, 0.478325, 0.270119, 0.158575, 0.213719, 0.265268

data shape = (5, 5, 5, 5), alpha=0.0001,beta=0.75,bias=1,size=3

[ce1adon]

I see your concern.
The lower and upper radius should be distinguished since they are different when size is even.
Moreover, the element corresponding to "end" should be included in the calculation. Current for-loop excluded it.

[ce1adon]

It looks like the current ctest config for lrn is not valid.
Current ctest chooses beta=0, which turned the test into a simple memory copy case without testing any features of lrn.
When beta is set to any non-zero value, the ctest failed.
Working on both using the correct verification method and modify kernels.

[ce1adon]

@scxiao
Hey Shucai, Please review the solution here:
https://github.com/AMDComputeLibraries/MLOpen/pull/2305

If you still encounter issues with the test, please let us know.

[daniellowell]

Fixed in 2.2.1