11 KiB
11 KiB
SEResNeXt与Res2Net系列
概述
ResNeXt是ResNet的典型变种网络之一，ResNeXt发表于2017年的CVPR会议。在此之前，提升模型精度的方法主要集中在将网络变深或者变宽，这样增加了参数量和计算量，推理速度也会相应变慢。ResNeXt结构提出了通道分组（cardinality）的概念，作者通过实验发现增加通道的组数比增加深度和宽度更有效。其可以在不增加参数复杂度的前提下提高准确率，同时还减少了参数的数量，所以是比较成功的ResNet的变种。
SENet是2017年ImageNet分类比赛的冠军方案，其提出了一个全新的SE结构，该结构可以迁移到任何其他网络中，其通过控制scale的大小，把每个通道间重要的特征增强，不重要的特征减弱，从而让提取的特征指向性更强。
Res2Net是2019年提出的一种全新的对ResNet的改进方案，该方案可以和现有其他优秀模块轻松整合，在不增加计算负载量的情况下，在ImageNet、CIFAR-100等数据集上的测试性能超过了ResNet。Res2Net结构简单，性能优越，进一步探索了CNN在更细粒度级别的多尺度表示能力。Res2Net揭示了一个新的提升模型精度的维度，即scale，其是除了深度、宽度和基数的现有维度之外另外一个必不可少的更有效的因素。该网络在其他视觉任务如目标检测、图像分割等也有相当不错的表现。
目前PaddleClas开源的这三类的预训练模型一共有24个，其指标如图所示，从图中可以看出，在同样Flops和Params下，改进版的模型往往有更高的精度，但是推理速度往往不如ResNet系列。另一方面，Res2Net表现也较为优秀，相比ResNeXt中的group操作、SEResNet中的SE结构操作，Res2Net在相同Flops、Params和推理速度下往往精度更佳。
注意：所有模型在预测时，图像的crop_size设置为224，resize_short_size设置为256。
精度、FLOPS和参数量
| Models | Top1 | Top5 | Reference top1 |
Reference top5 |
FLOPS (G) |
Parameters (M) |
|---|---|---|---|---|---|---|
| Res2Net50_26w_4s | 0.793 | 0.946 | 0.780 | 0.936 | 8.520 | 25.700 |
| Res2Net50_vd_26w_4s | 0.798 | 0.949 | 8.370 | 25.060 | ||
| Res2Net50_14w_8s | 0.795 | 0.947 | 0.781 | 0.939 | 9.010 | 25.720 |
| Res2Net101_vd_26w_4s | 0.806 | 0.952 | 16.670 | 45.220 | ||
| Res2Net200_vd_26w_4s | 0.812 | 0.957 | 31.490 | 76.210 | ||
| ResNeXt50_32x4d | 0.778 | 0.938 | 0.778 | 8.020 | 23.640 | |
| ResNeXt50_vd_32x4d | 0.796 | 0.946 | 8.500 | 23.660 | ||
| ResNeXt50_64x4d | 0.784 | 0.941 | 15.060 | 42.360 | ||
| ResNeXt50_vd_64x4d | 0.801 | 0.949 | 15.540 | 42.380 | ||
| ResNeXt101_32x4d | 0.787 | 0.942 | 0.788 | 15.010 | 41.540 | |
| ResNeXt101_vd_32x4d | 0.803 | 0.951 | 15.490 | 41.560 | ||
| ResNeXt101_64x4d | 0.784 | 0.945 | 0.796 | 29.050 | 78.120 | |
| ResNeXt101_vd_64x4d | 0.808 | 0.952 | 29.530 | 78.140 | ||
| ResNeXt152_32x4d | 0.790 | 0.943 | 22.010 | 56.280 | ||
| ResNeXt152_vd_32x4d | 0.807 | 0.952 | 22.490 | 56.300 | ||
| ResNeXt152_64x4d | 0.795 | 0.947 | 43.030 | 107.570 | ||
| ResNeXt152_vd_64x4d | 0.811 | 0.953 | 43.520 | 107.590 | ||
| SE_ResNet18_vd | 0.733 | 0.914 | 4.140 | 11.800 | ||
| SE_ResNet34_vd | 0.765 | 0.932 | 7.840 | 21.980 | ||
| SE_ResNet50_vd | 0.795 | 0.948 | 8.670 | 28.090 | ||
| SE_ResNeXt50_32x4d | 0.784 | 0.940 | 0.789 | 0.945 | 8.020 | 26.160 |
| SE_ResNeXt50_vd_32x4d | 0.802 | 0.949 | 10.760 | 26.280 | ||
| SE_ResNeXt101_32x4d | 0.791 | 0.942 | 0.793 | 0.950 | 15.020 | 46.280 |
| SENet154_vd | 0.814 | 0.955 | 45.830 | 114.290 |
FP16预测速度
| Models | batch_size=1 (ms) |
batch_size=4 (ms) |
batch_size=8 (ms) |
batch_size=32 (ms) |
|---|---|---|---|---|
| Res2Net50_26w_4s | 2.625 | 3.338 | 4.670 | 11.939 |
| Res2Net50_vd_26w_4s | 2.642 | 3.480 | 4.862 | 13.089 |
| Res2Net50_14w_8s | 3.393 | 4.237 | 5.473 | 13.979 |
| Res2Net101_vd_26w_4s | 5.128 | 6.190 | 7.995 | 20.534 |
| Res2Net200_vd_26w_4s | 9.594 | 11.131 | 14.278 | 36.258 |
| ResNeXt50_32x4d | 6.795 | 7.102 | 8.444 | 18.938 |
| ResNeXt50_vd_32x4d | 7.455 | 7.231 | 8.891 | 19.849 |
| ResNeXt50_64x4d | 20.279 | 12.343 | 13.633 | 32.772 |
| ResNeXt50_vd_64x4d | 16.325 | 21.773 | 25.007 | 55.329 |
| ResNeXt101_32x4d | 14.847 | 15.092 | 15.847 | 42.681 |
| ResNeXt101_vd_32x4d | 15.227 | 15.139 | 16.603 | 39.371 |
| ResNeXt101_64x4d | 28.221 | 29.455 | 29.873 | 59.415 |
| ResNeXt101_vd_64x4d | 31.051 | 28.160 | 28.915 | 60.525 |
| ResNeXt152_32x4d | 22.961 | 23.167 | 24.173 | 51.621 |
| ResNeXt152_vd_32x4d | 23.259 | 23.469 | 23.886 | 52.085 |
| ResNeXt152_64x4d | 41.930 | 42.441 | 45.985 | 79.405 |
| ResNeXt152_vd_64x4d | 42.778 | 43.281 | 45.017 | 79.728 |
| SE_ResNet18_vd | 1.256 | 1.463 | 1.917 | 4.316 |
| SE_ResNet34_vd | 2.314 | 2.691 | 3.432 | 7.411 |
| SE_ResNet50_vd | 2.884 | 4.051 | 5.421 | 15.013 |
| SE_ResNeXt50_32x4d | 7.973 | 10.613 | 12.788 | 29.091 |
| SE_ResNeXt50_vd_32x4d | 8.340 | 12.245 | 15.253 | 30.399 |
| SE_ResNeXt101_32x4d | 17.324 | 21.004 | 28.541 | 52.888 |
| SENet154_vd | 47.234 | 48.018 | 52.967 | 109.787 |
FP32预测速度
| Models | batch_size=1 (ms) |
batch_size=4 (ms) |
batch_size=8 (ms) |
batch_size=32 (ms) |
|---|---|---|---|---|
| Res2Net50_26w_4s | 3.711 | 5.855 | 8.450 | 26.084 |
| Res2Net50_vd_26w_4s | 3.651 | 5.986 | 8.747 | 26.772 |
| Res2Net50_14w_8s | 4.549 | 6.863 | 9.492 | 27.049 |
| Res2Net101_vd_26w_4s | 6.658 | 10.870 | 15.364 | 47.054 |
| Res2Net200_vd_26w_4s | 12.017 | 19.871 | 28.330 | 88.645 |
| ResNeXt50_32x4d | 6.747 | 8.862 | 11.961 | 32.782 |
| ResNeXt50_vd_32x4d | 6.746 | 9.037 | 12.279 | 33.496 |
| ResNeXt50_64x4d | 11.577 | 14.570 | 20.425 | 57.979 |
| ResNeXt50_vd_64x4d | 19.219 | 21.454 | 30.943 | 90.950 |
| ResNeXt101_32x4d | 14.652 | 18.082 | 24.148 | 70.200 |
| ResNeXt101_vd_32x4d | 14.927 | 18.454 | 23.894 | 67.334 |
| ResNeXt101_64x4d | 28.726 | 30.999 | 43.169 | 116.282 |
| ResNeXt101_vd_64x4d | 28.350 | 31.186 | 41.315 | 113.655 |
| ResNeXt152_32x4d | 23.578 | 27.323 | 35.588 | 99.121 |
| ResNeXt152_vd_32x4d | 23.548 | 26.879 | 35.091 | 104.832 |
| ResNeXt152_64x4d | 43.214 | 43.339 | 60.990 | 159.381 |
| ResNeXt152_vd_64x4d | 43.998 | 44.510 | 61.094 | 160.601 |
| SE_ResNet18_vd | 1.353 | 1.867 | 3.021 | 9.331 |
| SE_ResNet34_vd | 2.421 | 3.201 | 5.294 | 16.849 |
| SE_ResNet50_vd | 3.403 | 6.023 | 8.721 | 26.978 |
| SE_ResNeXt50_32x4d | 8.339 | 12.689 | 15.471 | 41.562 |
| SE_ResNeXt50_vd_32x4d | 7.849 | 13.530 | 16.810 | 44.020 |
| SE_ResNeXt101_32x4d | 16.853 | 24.409 | 32.666 | 81.806 |
| SENet154_vd | 46.002 | 53.666 | 70.589 | 180.334 |