Support autoenvolved anchors as ultralytics

[Root970103]

When converting the yolov5 custom model to yolort, the configuration of Anchors is fixed. This will cause a large gap between the two prediction results.

This is my test results for reference.

Detection results with ultralytics

tensor([[ 60.20518, 275.15906, 227.31180, 460.78314,   0.95166,   2.00000],
        [239.06186,  78.73691, 411.81058, 264.97150,   0.95022,   0.00000],
        [245.05939, 421.12323, 407.59259, 505.68793,   0.94337,   2.00000],
        [241.37396, 348.25543, 413.31641, 406.12311,   0.93951,   0.00000],
        [ 68.77753, 465.43747, 217.82889, 519.14197,   0.93248,   5.00000],
        [245.88387, 509.54541, 404.57898, 541.54272,   0.92447,   5.00000],
        [ 60.48768,  78.21830, 225.84097, 269.68933,   0.91452,   0.00000],
        [242.17107, 550.16064, 413.19885, 585.58118,   0.90575,   0.00000],
        [ 60.04885, 564.58539, 184.14281, 573.51935,   0.88148,   6.00000],
        [257.21490, 295.98309, 397.26947, 315.52979,   0.87154,   7.00000],
        [270.75903, 335.44797, 379.67609, 346.24161,   0.86142,   7.00000],
        [276.42148, 269.73453, 380.06650, 281.44186,   0.85676,   7.00000],
        [ 58.77311, 539.35565, 129.32153, 550.79681,   0.84660,   6.00000],
        [299.25656, 317.71469, 353.17801, 333.91336,   0.84569,   7.00000],
        [ 70.90355, 576.13129, 228.42677, 586.76593,   0.82126,   0.00000],
        [272.83978, 282.16409, 379.86005, 293.18692,   0.67504,   7.00000]])

Detection results with yolort

boxes:

tensor([[114.73997, 286.17834, 172.77701, 449.76385],
        [296.58661, 128.59970, 354.28583, 215.10870],
        [298.10165, 426.14325, 354.55032, 500.66791],
        [310.67633, -24.03447, 340.96069, 367.68439],
        [298.63022, 363.74905, 356.06015, 390.62949],
        [310.64130, 375.32629, 340.15778, 551.87183],
        [279.98190,  71.63794, 372.35941, 273.37885],
        [281.49362, 345.05667, 372.36020, 408.86185],
        [117.42003, 468.62558, 169.18639, 515.95392],
        [318.71570, 317.67377, 335.81323, 436.61597],
        [226.89404,  79.70534, 424.16278, 265.10114],
        [ 46.96851, 273.46609, 239.71344, 461.19077],
        [315.01773, 509.35077, 335.44510, 541.73737],
        [302.44864, 511.76874, 347.06296, 539.38751],
        [114.45029,  89.58469, 171.87836, 258.32294],
        [317.77310, 491.92383, 333.15158, 558.55872],
        [299.12274, 559.64423, 356.24719, 576.09760],
        [318.85751, 531.31708, 335.88858, 604.20856],
        [135.67125, 436.88367, 151.06697, 547.62537],
        [ 44.50909,  77.33535, 241.77551, 270.65427],
        [114.10909, 564.53101, 130.08257, 573.57373],
        [128.94606, 174.29877, 157.97704, 560.87738],
        [318.22821, 295.86420, 336.25616, 315.64868],
        [127.40814, -24.54239, 158.43588, 373.00723],
        [231.56476, 419.62057, 420.35730, 506.79044],
        [105.16776, 565.24457, 139.56566, 572.73053],
        [319.92688, 283.30414, 333.55621, 328.24921],
        [316.60138, 550.55896, 338.71069, 585.45984],
        [318.20761, 335.38226, 332.22751, 346.30725],
        [307.69409, 297.45999, 346.46869, 314.14407],
        [310.38986, 336.16382, 340.24734, 345.19263],
        [321.57333, 269.66330, 334.91464, 281.51309],
        [116.27255, 558.97729, 128.35686, 579.16809],
        [319.82718, 328.34348, 330.08755, 353.56308],
        [ 89.50679, 539.28607,  98.58784, 550.86639],
        [322.74689, 317.61615, 329.68768, 334.01190],
        [313.31296, 270.62665, 342.22092, 280.61163],
        [318.55316, 318.89578, 333.70508, 332.48337],
        [323.54080, 307.02875, 328.81540, 344.24860],
        [ 84.36114, 540.25403, 104.05352, 550.03149],
        [323.00452, 262.09003, 333.10120, 289.10217],
        [ 90.60516, 532.02325,  97.54469, 558.19513],
        [127.85805, 576.88733, 171.47224, 586.00989],
        [141.52585, 569.54669, 157.04645, 593.17780],
        [139.58984, 575.94946, 159.91479, 586.71265],
        [319.46204, 282.09702, 333.23779, 293.25394],
        [311.79651, 283.07559, 341.74689, 292.15579],
        [321.21085, 275.04596, 331.59256, 300.28229],
        [322.56943, 287.68039, 336.84097, 324.22037]])

scores:

tensor([0.95166, 0.95022, 0.94337, 0.94313, 0.93951, 0.93825, 0.93778, 0.93444, 0.93248, 0.93215, 0.92636, 0.92569, 0.92447, 0.92406, 0.91452, 0.90793, 0.90575, 0.90452, 0.89434, 0.89394, 0.88148, 0.8779
0, 0.87154, 0.86893, 0.86798, 0.86740, 0.86400, 0.86210, 0.86142, 0.86071, 0.85727, 0.85676, 0.85073, 0.85071,
        0.84660, 0.84569, 0.84113, 0.83803, 0.83582, 0.83308, 0.82758, 0.82343, 0.82126, 0.79421, 0.78466, 0.67504, 0.66505, 0.66226, 0.28443])

labels:

tensor([2, 0, 2, 0, 0, 2, 0, 0, 5, 0, 0, 2, 5, 5, 0, 5, 0, 0, 5, 0, 6, 2, 7, 0, 2, 6, 7, 0, 7, 7, 7, 7, 6, 7, 6, 7, 7, 7, 7, 6, 7, 6, 0, 0, 0, 7, 7, 7, 0])

I also converted the official model, and there is no difference between the two models. So I hope someone can provide the method to change the default configuration of Anchors.

[Tomakko]

I just fixed this issue today, I will make a PR for this in the next days!

[zhiqwang]

I just fixed this issue today, I will make a PR for this in the next days!

Thanks @Tomakko , we are eager for your contribution!

[zhiqwang]

FYI, Now we can attach the anchor_grids from ultralytics as following:

model = torch.hub.load('ultralytics/yolov5', 'custom', path=ultralytics_weights_path)
anchor_grids = model.model.model[-1].anchor_grid.view(3, -1).cpu().numpy().tolist()

And then pass it into yolort:

from yolort.models.yolo import yolov5_darknet_pan_s_r40 as yolov5s
model = yolov5s(anchor_grids=anchor_grids, num_classes=10)

[Tomakko]

Make sure to disable autoshape in
model = torch.hub.load('ultralytics/yolov5', 'custom', path=ultralytics_weights_path, autoshape=False)