@@ -42,7 +42,62 @@ The original Caffe implementation can be found here : [Original Caffe implementa
4242
4343SimpleNet performs very decently, it outperforms VGGNet, variants of ResNet and MobileNets(1-3)
4444and its pretty fast as well! and its all using plain old CNN!.
45- For benchmark results [ look here] ( https://github.com/Coderx7/SimpleNet_Pytorch/tree/master/ImageNet/training_scripts/imagenet_training/results )
45+ 46+ Here's an example of benchmark run on small variants of simplenet and some other known architectures such as mobilenets.
47+ Small variants of simplenet consistently achieve high performance/accuracy:
48+ 49+ | model | samples_per_sec | param_count | top1 | top5 |
50+ | :----------------------------------| :--------------: | :-----------: | :--: | :---: |
51+ | mobilenetv3_small_050 | 3035.37 | 1.59 | 57.89 | 80.194 |
52+ | ** simplenetv1_small_m1_05** | 2839.35 | 1.51 | ** 60.89** | ** 82.978** |
53+ | lcnet_050 | 2683.57 | 1.88 | 63.1 | 84.382 |
54+ | ** simplenetv1_small_m2_05** | 2340.51 | 1.51 | ** 61.524** | ** 83.432** |
55+ | mobilenetv3_small_075 | 1781.14 | 2.04 | 65.242 | 85.438 |
56+ | tf_mobilenetv3_small_075 | 1674.31 | 2.04 | 65.714 | 86.134 |
57+ | ** simplenetv1_small_m1_075** | 1524.64 | 3.29 | ** 67.764** | ** 87.66** |
58+ | tf_mobilenetv3_small_minimal_100 | 1308.27 | 2.04 | 62.908 | 84.234 |
59+ | ** simplenetv1_small_m2_075** | 1264.33 | 3.29 | ** 68.15** | ** 87.762** |
60+ | mobilenetv3_small_100 | 1263.23 | 2.54 | 67.656 | 87.634 |
61+ | tf_mobilenetv3_small_100 | 1220.08 | 2.54 | 67.924 | 87.664 |
62+ | mnasnet_small | 1085.15 | 2.03 | 66.206 | 86.508 |
63+ | mobilenetv2_050 | 848.38 | 1.97 | 65.942 | 86.082 |
64+ | dla46_c | 531.0 | 1.3 | 64.866 | 86.294 |
65+ | dla46x_c | 318.32 | 1.07 | 65.97 | 86.98 |
66+ | dla60x_c | 298.59 | 1.32 | 67.892 | 88.426 |
67+ 68+ and this is a sample for larger models: simplenet variants outperform many newer architecures.
69+ 70+ | model | samples_per_sec | param_count | top1 | top5 |
71+ | :----------------------------------| :--------------: | :-----------: | :--: | :---: |
72+ | vit_tiny_r_s16_p8_224 | 1882.23 | 6.34 | 71.792 | 90.822 |
73+ | simplenetv1_small_m1_075 | 1516.74 | 3.29 | 67.764 | 87.660 |
74+ | simplenetv1_small_m2_075 | 1260.89 | 3.29 | 68.150 | 87.762 |
75+ | simplenetv1_5m_m1 | 1107.70 | 5.75 | 71.370 | 90.100 |
76+ | deit_tiny_patch16_224 | 991.41 | 5.72 | 72.172 | 91.114 |
77+ | resnet18 | 876.92 | 11.69 | 69.744 | 89.082 |
78+ | simplenetv1_5m_m2 | 835.17 | 5.75 | 71.936 | 90.300 |
79+ | crossvit_9_240 | 602.13 | 8.55 | 73.960 | 91.968 |
80+ | vit_base_patch32_224_sam | 571.37 | 88.22 | 73.694 | 91.010 |
81+ | tinynet_b | 530.15 | 3.73 | 74.976 | 92.184 |
82+ | resnet26 | 524.36 | 16.00 | 75.300 | 92.578 |
83+ | tf_mobilenetv3_large_075 | 505.13 | 3.99 | 73.436 | 91.344 |
84+ | resnet34 | 491.96 | 21.80 | 75.114 | 92.284 |
85+ | regnetx_006 | 478.41 | 6.20 | 73.860 | 91.672 |
86+ | dla34 | 472.49 | 15.74 | 74.620 | 92.072 |
87+ | simplenetv1_9m_m1 | 459.21 | 9.51 | 73.376 | 91.048 |
88+ | repvgg_b0 | 455.36 | 15.82 | 75.160 | 92.418 |
89+ | ghostnet_100 | 407.03 | 5.18 | 73.974 | 91.460 |
90+ | tf_mobilenetv3_large_minimal_100 | 406.84 | 3.92 | 72.250 | 90.630 |
91+ | mobilenetv3_large_100 | 402.08 | 5.48 | 75.766 | 92.544 |
92+ | simplenetv1_9m_m2 | 389.94 | 9.51 | 74.170 | 91.614 |
93+ | tf_mobilenetv3_large_100 | 388.30 | 5.48 | 75.518 | 92.604 |
94+ | mobilenetv2_100 | 295.68 | 3.50 | 72.970 | 91.020 |
95+ | densenet121 | 293.94 | 7.98 | 75.584 | 92.652 |
96+ | mnasnet_100 | 262.25 | 4.38 | 74.658 | 92.112 |
97+ 98+ 99+ Benchmark was done using a GTX1080 on Pytorch 1.11 with fp32, nhwc, batchsize of 256, input size = ` 224x224x3 ` .
100+ For all benchmark results [ look here] ( https://github.com/Coderx7/SimpleNet_Pytorch/tree/master/ImageNet/training_scripts/imagenet_training/results )
46101
47102-- The models pretrained weights (pytorch, onnx, jit) can be found in [ Release section] ( https://github.com/Coderx7/SimpleNet_Pytorch/releases )
48103
0 commit comments