diff --git a/configs/pretrained.list b/configs/pretrained.list index a3ef2e9335fd9ce032afda5172c6a06499cc3d86..b01272ea2bc7ed0d9587f486001a1bdb3d7697d9 100644 --- a/configs/pretrained.list +++ b/configs/pretrained.list @@ -12,6 +12,7 @@ ResNet101_vd ResNet152_vd ResNet200_vd ResNet50_vd_ssld +ResNet101_vd_ssld MobileNetV3_large_x0_35 MobileNetV3_large_x0_5 MobileNetV3_large_x0_75 diff --git a/docs/images/models/DPN.png b/docs/images/models/DPN.png deleted file mode 100644 index f84e9c22e6e57c2afa810aab935afe835d0eb63b..0000000000000000000000000000000000000000 Binary files a/docs/images/models/DPN.png and /dev/null differ diff --git a/docs/images/models/DPN.png.flops.png b/docs/images/models/DPN.png.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..72bb96f49812711035ec09ce0d8d44202d17cfcb Binary files /dev/null and b/docs/images/models/DPN.png.flops.png differ diff --git a/docs/images/models/DPN.png.fp32.png b/docs/images/models/DPN.png.fp32.png new file mode 100644 index 0000000000000000000000000000000000000000..bcd524c781d305fdacfd5a07886851fadbeeeddc Binary files /dev/null and b/docs/images/models/DPN.png.fp32.png differ diff --git a/docs/images/models/DPN.png.params.png b/docs/images/models/DPN.png.params.png new file mode 100644 index 0000000000000000000000000000000000000000..818f1961bbaf98cd0aff0c5405113f107fa02e54 Binary files /dev/null and b/docs/images/models/DPN.png.params.png differ diff --git a/docs/images/models/EfficientNet.png.flops.png b/docs/images/models/EfficientNet.png.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..dd3c36ced2973133bdd0bb6b300125b25fdcefe4 Binary files /dev/null and b/docs/images/models/EfficientNet.png.flops.png differ diff --git a/docs/images/models/EfficientNet.png.fp32.png b/docs/images/models/EfficientNet.png.fp32.png new file mode 100644 index 0000000000000000000000000000000000000000..eca753f7d84699cebee13d15484291d96f0a9b6f Binary files /dev/null and b/docs/images/models/EfficientNet.png.fp32.png differ diff --git a/docs/images/models/EfficientNet.png.params.png b/docs/images/models/EfficientNet.png.params.png new file mode 100644 index 0000000000000000000000000000000000000000..2348c55013998bcb2ed2c3edff282493243ee37a Binary files /dev/null and b/docs/images/models/EfficientNet.png.params.png differ diff --git a/docs/images/models/HRNet.png b/docs/images/models/HRNet.png deleted file mode 100644 index 124fd3d43a25d085122e5acab57c8e78c9f32fa4..0000000000000000000000000000000000000000 Binary files a/docs/images/models/HRNet.png and /dev/null differ diff --git a/docs/images/models/HRNet.png.flops.png b/docs/images/models/HRNet.png.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..5f8ce9cd2c1c8ed9e8fb775bd89a8acd3a8e9402 Binary files /dev/null and b/docs/images/models/HRNet.png.flops.png differ diff --git a/docs/images/models/HRNet.png.fp32.png b/docs/images/models/HRNet.png.fp32.png new file mode 100644 index 0000000000000000000000000000000000000000..0e73fb4b57dc374560efa92429a2dd457c73369e Binary files /dev/null and b/docs/images/models/HRNet.png.fp32.png differ diff --git a/docs/images/models/HRNet.png.params.png b/docs/images/models/HRNet.png.params.png new file mode 100644 index 0000000000000000000000000000000000000000..e4443a770ac0ca910d1158fe9adaf6dd92e680aa Binary files /dev/null and b/docs/images/models/HRNet.png.params.png differ diff --git a/docs/images/models/Inception.png b/docs/images/models/Inception.png deleted file mode 100644 index 50c08ce643bc45ed7054e06bacf52d25c4cd49b0..0000000000000000000000000000000000000000 Binary files a/docs/images/models/Inception.png and /dev/null differ diff --git a/docs/images/models/Inception.png.flops.png b/docs/images/models/Inception.png.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..589f3931c1feef3e1c0245566cd0c7e0a22782d8 Binary files /dev/null and b/docs/images/models/Inception.png.flops.png differ diff --git a/docs/images/models/Inception.png.fp32.png b/docs/images/models/Inception.png.fp32.png new file mode 100644 index 0000000000000000000000000000000000000000..b9245800a2d7ca6fad5ed7457e55356d579a81d0 Binary files /dev/null and b/docs/images/models/Inception.png.fp32.png differ diff --git a/docs/images/models/Inception.png.params.png b/docs/images/models/Inception.png.params.png new file mode 100644 index 0000000000000000000000000000000000000000..657c4360451b24905b9a1ce170e4da719d35d917 Binary files /dev/null and b/docs/images/models/Inception.png.params.png differ diff --git a/docs/images/models/ResNet.png b/docs/images/models/ResNet.png deleted file mode 100644 index fb992c694da4b940ca2dea7f584fa3dc179fe35e..0000000000000000000000000000000000000000 Binary files a/docs/images/models/ResNet.png and /dev/null differ diff --git a/docs/images/models/ResNet.png.flops.png b/docs/images/models/ResNet.png.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..da1fd2eb359f57fbd545a5436b737fe23df6e891 Binary files /dev/null and b/docs/images/models/ResNet.png.flops.png differ diff --git a/docs/images/models/ResNet.png.fp32.png b/docs/images/models/ResNet.png.fp32.png new file mode 100644 index 0000000000000000000000000000000000000000..05020997f2b6eb2c926d2a8948ed69b393b6cd3b Binary files /dev/null and b/docs/images/models/ResNet.png.fp32.png differ diff --git a/docs/images/models/ResNet.png.params.png b/docs/images/models/ResNet.png.params.png new file mode 100644 index 0000000000000000000000000000000000000000..6fcbb69cc1e1e9a3402f2849fe016a73312d9a79 Binary files /dev/null and b/docs/images/models/ResNet.png.params.png differ diff --git a/docs/images/models/SeResNeXt.png b/docs/images/models/SeResNeXt.png deleted file mode 100644 index 086a0c0c1018a20b2fd7adef4f82566aa54e4c1c..0000000000000000000000000000000000000000 Binary files a/docs/images/models/SeResNeXt.png and /dev/null differ diff --git a/docs/images/models/SeResNeXt.png.flops.png b/docs/images/models/SeResNeXt.png.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..51d6d6497e9cd582ad671a79b9e24bbdc1a9bdae Binary files /dev/null and b/docs/images/models/SeResNeXt.png.flops.png differ diff --git a/docs/images/models/SeResNeXt.png.fp32.png b/docs/images/models/SeResNeXt.png.fp32.png new file mode 100644 index 0000000000000000000000000000000000000000..452488955096f896ecb9dafe07885f666c92d8ad Binary files /dev/null and b/docs/images/models/SeResNeXt.png.fp32.png differ diff --git a/docs/images/models/SeResNeXt.png.params.png b/docs/images/models/SeResNeXt.png.params.png new file mode 100644 index 0000000000000000000000000000000000000000..9898f52fb0be6bdfb22d39a1f4a16c98e20ab510 Binary files /dev/null and b/docs/images/models/SeResNeXt.png.params.png differ diff --git a/docs/images/models/main_fps_top1.png b/docs/images/models/main_fps_top1.png old mode 100644 new mode 100755 index e27b8bb9d80273c0d494d563e747a28f3a1f37eb..16be5ba63350943676a617fbb6f982f807b274f8 Binary files a/docs/images/models/main_fps_top1.png and b/docs/images/models/main_fps_top1.png differ diff --git a/docs/images/models/mobile_arm_storage.png b/docs/images/models/mobile_arm_storage.png old mode 100644 new mode 100755 diff --git a/docs/images/models/mobile_arm_top1.png b/docs/images/models/mobile_arm_top1.png old mode 100644 new mode 100755 diff --git a/docs/images/models/mobile_trt.png.flops.png b/docs/images/models/mobile_trt.png.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..6e7010906c58ef9e1f4f286725442ca85e638c51 Binary files /dev/null and b/docs/images/models/mobile_trt.png.flops.png differ diff --git a/docs/images/models/mobile_trt.png.params.png b/docs/images/models/mobile_trt.png.params.png new file mode 100644 index 0000000000000000000000000000000000000000..35b78a65b390f9843def8870bbd63f656070cdcc Binary files /dev/null and b/docs/images/models/mobile_trt.png.params.png differ diff --git a/docs/zh_CN/models/DPN_DenseNet.md b/docs/zh_CN/models/DPN_DenseNet.md index 3c4e6e694dfeba326d2df2dcd7f71459c1a8c4aa..4dc811bac54b3e01a606b8039ec9882374888bbe 100644 --- a/docs/zh_CN/models/DPN_DenseNet.md +++ b/docs/zh_CN/models/DPN_DenseNet.md @@ -2,9 +2,16 @@ ## 概述 正在持续更新中...... -![](../../images/models/DPN.png) -所有模型在预测时，图像的crop_size设置为224，resize_short_size设置为256。 +该系列模型的FLOPS、参数量以及fp32预测耗时如下图所示。 + +![](../../images/models/DPN.png.flops.png) + +![](../../images/models/DPN.png.params.png) + +![](../../images/models/DPN.png.fp32.png) + +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 @@ -22,33 +29,19 @@ | DPN131 | 0.807 | 0.951 | 0.801 | 0.949 | 30.510 | 75.360 | -## FP16预测速度 - -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| DenseNet121 | 3.653 | 4.560 | 5.574 | 11.517 | -| DenseNet161 | 7.826 | 8.936 | 10.970 | 22.554 | -| DenseNet169 | 5.625 | 6.698 | 7.876 | 14.983 | -| DenseNet201 | 7.243 | 8.537 | 10.111 | 18.928 | -| DenseNet264 | 10.882 | 12.539 | 14.645 | 26.455 | -| DPN68 | 10.310 | 11.060 | 14.299 | 29.618 | -| DPN92 | 16.335 | 17.373 | 23.197 | 45.210 | -| DPN98 | 18.975 | 23.073 | 28.902 | 66.280 | -| DPN107 | 24.932 | 28.607 | 37.513 | 89.112 | -| DPN131 | 25.425 | 29.874 | 37.355 | 88.583 | ## FP32预测速度 -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| DenseNet121 | 3.732 | 6.614 | 8.517 | 21.755 | -| DenseNet161 | 8.282 | 14.438 | 19.336 | 51.953 | -| DenseNet169 | 5.705 | 10.074 | 12.432 | 28.756 | -| DenseNet201 | 7.315 | 13.830 | 16.941 | 38.654 | -| DenseNet264 | 10.986 | 21.460 | 25.724 | 56.501 | -| DPN68 | 10.357 | 11.025 | 14.903 | 34.380 | -| DPN92 | 16.067 | 21.315 | 26.176 | 62.126 | -| DPN98 | 18.455 | 26.710 | 36.009 | 104.084 | -| DPN107 | 24.164 | 37.691 | 51.307 | 148.041 | -| DPN131 | 24.761 | 35.806 | 48.401 | 133.233 | +| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +|-------------|-----------|-------------------|--------------------------| +| DenseNet121 | 224 | 256 | 4.371 | +| DenseNet161 | 224 | 256 | 8.863 | +| DenseNet169 | 224 | 256 | 6.391 | +| DenseNet201 | 224 | 256 | 8.173 | +| DenseNet264 | 224 | 256 | 11.942 | +| DPN68 | 224 | 256 | 11.805 | +| DPN92 | 224 | 256 | 17.840 | +| DPN98 | 224 | 256 | 21.057 | +| DPN107 | 224 | 256 | 28.685 | +| DPN131 | 224 | 256 | 28.083 | diff --git a/docs/zh_CN/models/EfficientNet_and_ResNeXt101_wsl.md b/docs/zh_CN/models/EfficientNet_and_ResNeXt101_wsl.md index f4f790abdac098067577150da60ef6ee2f0b4629..165d85367a577f5f84d4bb2bcb350266300e4baa 100644 --- a/docs/zh_CN/models/EfficientNet_and_ResNeXt101_wsl.md +++ b/docs/zh_CN/models/EfficientNet_and_ResNeXt101_wsl.md @@ -2,26 +2,14 @@ ## 概述 正在持续更新中...... -![](../../images/models/EfficientNet.png) -在预测时，图像的crop_size和resize_short_size如下表所示。 +该系列模型的FLOPS、参数量以及fp32预测耗时如下图所示。 -| Models | crop_size | resize_short_size | -|:--:|:--:|:--:| -| ResNeXt101_32x8d_wsl | 224 | 224 | -| ResNeXt101_32x16d_wsl | 224 | 224 | -| ResNeXt101_32x32d_wsl | 224 | 224 | -| ResNeXt101_32x48d_wsl | 224 | 224 | -| Fix_ResNeXt101_32x48d_wsl | 320 | 320 | -| EfficientNetB0 | 224 | 256 | -| EfficientNetB1 | 240 | 272 | -| EfficientNetB2 | 260 | 292 | -| EfficientNetB3 | 300 | 332 | -| EfficientNetB4 | 380 | 412 | -| EfficientNetB5 | 456 | 488 | -| EfficientNetB6 | 528 | 560 | -| EfficientNetB7 | 600 | 632 | -| EfficientNetB0_small | 224 | 256 | +![](../../images/models/EfficientNet.png.flops.png) + +![](../../images/models/EfficientNet.png.params.png) + +![](../../images/models/EfficientNet.png.fp32.png) ## 精度、FLOPS和参数量 @@ -44,41 +32,21 @@ | EfficientNetB0_
small | 0.758 | 0.926 | | | 0.720 | 4.650 | -## FP16预测速度 - -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| ResNeXt101_
32x8d_wsl | 16.063 | 16.342 | 24.914 | 45.035 | -| ResNeXt101_
32x16d_wsl | 16.471 | 25.235 | 30.762 | 67.869 | -| ResNeXt101_
32x32d_wsl | 29.425 | 37.149 | 50.834 | | -| ResNeXt101_
32x48d_wsl | 40.311 | 58.414 | | | -| Fix_ResNeXt101_
32x48d_wsl | 43.960 | 86.514 | | | -| EfficientNetB0 | 1.759 | 2.748 | 3.761 | 10.178 | -| EfficientNetB1 | 2.592 | 4.122 | 5.829 | 16.262 | -| EfficientNetB2 | 2.866 | 4.715 | 7.064 | 20.954 | -| EfficientNetB3 | 3.869 | 6.815 | 10.672 | 34.097 | -| EfficientNetB4 | 5.626 | 11.937 | 19.753 | 67.436 | -| EfficientNetB5 | 8.907 | 21.685 | 37.248 | 134.185 | -| EfficientNetB6 | 13.591 | 34.093 | 60.976 | | -| EfficientNetB7 | 20.963 | 56.397 | 103.971 | | -| EfficientNetB0_
small | 1.039 | 1.665 | 2.493 | 7.748 | - - ## FP32预测速度 -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| ResNeXt101_
32x8d_wsl | 16.325 | 25.633 | 37.196 | 108.535 | -| ResNeXt101_
32x16d_wsl | 25.224 | 40.929 | 62.898 | | -| ResNeXt101_
32x32d_wsl | 41.047 | 79.575 | | | -| ResNeXt101_
32x48d_wsl | 60.610 | | | | -| Fix_ResNeXt101_
32x48d_wsl | 80.280 | | | | -| EfficientNetB0 | 1.902 | 3.296 | 4.361 | 11.319 | -| EfficientNetB1 | 2.908 | 5.093 | 6.900 | 18.015 | -| EfficientNetB2 | 3.324 | 5.832 | 8.357 | 23.371 | -| EfficientNetB3 | 4.557 | 8.526 | 12.485 | 38.124 | -| EfficientNetB4 | 6.767 | 14.742 | 23.218 | 77.590 | -| EfficientNetB5 | 11.097 | 26.642 | 43.590 | | -| EfficientNetB6 | 17.582 | 42.408 | 74.336 | | -| EfficientNetB7 | 26.529 | 70.337 | 126.839 | | -| EfficientNetB0_
small | 1.171 | 2.026 | 2.906 | 8.506 | +| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +|-------------------------------|-----------|-------------------|--------------------------| +| ResNeXt101_
32x8d_wsl | 224 | 256 | 19.127 | +| ResNeXt101_
32x16d_wsl | 224 | 256 | 23.629 | +| ResNeXt101_
32x32d_wsl | 224 | 256 | 40.214 | +| ResNeXt101_
32x48d_wsl | 224 | 256 | 59.714 | +| Fix_ResNeXt101_
32x48d_wsl | 320 | 320 | 82.431 | +| EfficientNetB0 | 224 | 256 | 2.449 | +| EfficientNetB1 | 240 | 272 | 3.547 | +| EfficientNetB2 | 260 | 292 | 3.908 | +| EfficientNetB3 | 300 | 332 | 5.145 | +| EfficientNetB4 | 380 | 412 | 7.609 | +| EfficientNetB5 | 456 | 488 | 12.078 | +| EfficientNetB6 | 528 | 560 | 18.381 | +| EfficientNetB7 | 600 | 632 | 27.817 | +| EfficientNetB0_
small | 224 | 256 | 1.692 | diff --git a/docs/zh_CN/models/HRNet.md b/docs/zh_CN/models/HRNet.md index ac42a6694b2d8ae0572294a6fcd5fc60001804e0..576bc301cf51f0652e015329626b36fccf6f0ea9 100644 --- a/docs/zh_CN/models/HRNet.md +++ b/docs/zh_CN/models/HRNet.md @@ -2,8 +2,14 @@ ## 概述 正在持续更新中...... -![](../../images/models/HRNet.png) -所有模型在预测时，图像的crop_size设置为224，resize_short_size设置为256。 + +该系列模型的FLOPS、参数量以及fp32预测耗时如下图所示。 + +![](../../images/models/HRNet.png.flops.png) + +![](../../images/models/HRNet.png.params.png) + +![](../../images/models/HRNet.png.fp32.png) ## 精度、FLOPS和参数量 @@ -19,27 +25,14 @@ | HRNet_W64_C | 0.793 | 0.946 | 0.795 | 0.946 | 57.830 | 128.060 | -## FP16预测速度 - -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| HRNet_W18_C | 6.188 | 7.207 | 9.149 | 18.221 | -| HRNet_W30_C | 7.941 | 8.851 | 10.540 | 21.129 | -| HRNet_W32_C | 7.904 | 8.890 | 10.752 | 21.159 | -| HRNet_W40_C | 9.233 | 11.600 | 13.927 | 29.868 | -| HRNet_W44_C | 9.917 | 12.119 | 15.555 | 31.948 | -| HRNet_W48_C | 10.198 | 12.399 | 15.572 | 32.210 | -| HRNet_W64_C | 12.264 | 14.552 | 18.251 | 41.106 | - - ## FP32预测速度 -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| HRNet_W18_C | 6.828 | 8.552 | 11.154 | 30.665 | -| HRNet_W30_C | 8.901 | 11.067 | 14.421 | 43.459 | -| HRNet_W32_C | 8.983 | 11.334 | 14.688 | 44.564 | -| HRNet_W40_C | 10.300 | 14.720 | 20.257 | 64.346 | -| HRNet_W44_C | 11.183 | 15.830 | 25.292 | 73.136 | -| HRNet_W48_C | 11.619 | 16.791 | 26.569 | 77.536 | -| HRNet_W64_C | 14.434 | 20.988 | 35.114 | 117.434 | +| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +|-------------|-----------|-------------------|--------------------------| +| HRNet_W18_C | 224 | 256 | 7.368 | +| HRNet_W30_C | 224 | 256 | 9.402 | +| HRNet_W32_C | 224 | 256 | 9.467 | +| HRNet_W40_C | 224 | 256 | 10.739 | +| HRNet_W44_C | 224 | 256 | 11.497 | +| HRNet_W48_C | 224 | 256 | 12.165 | +| HRNet_W64_C | 224 | 256 | 15.003 | diff --git a/docs/zh_CN/models/Inception.md b/docs/zh_CN/models/Inception.md index f23defe2743ea736203545567adc23c599cd5db7..61d168b2dd8bf5065c1267d30bd29c32cf62316a 100644 --- a/docs/zh_CN/models/Inception.md +++ b/docs/zh_CN/models/Inception.md @@ -2,8 +2,14 @@ ## 概述 正在持续更新中...... -![](../../images/models/Inception.png) -GoogLeNet在预测时，图像的crop_size设置为224，resize_short_size设置为256，其余模型在预测时，图像的crop_size设置为299，resize_short_size设置为320。 + +该系列模型的FLOPS、参数量以及fp32预测耗时如下图所示。 + +![](../../images/models/Inception.png.flops.png) + +![](../../images/models/Inception.png.params.png) + +![](../../images/models/Inception.png.fp32.png) ## 精度、FLOPS和参数量 @@ -19,27 +25,15 @@ GoogLeNet在预测时，图像的crop_size设置为224，resize_short_size设置 | InceptionV4 | 0.808 | 0.953 | 0.800 | 0.950 | 24.570 | 42.680 | -## FP16预测速度 - -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| GoogLeNet | 1.428 | 1.833 | 2.138 | 4.143 | -| Xception41 | 1.545 | 2.772 | 4.961 | 18.447 | -| Xception41
_deeplab | 1.630 | 2.647 | 4.462 | 16.354 | -| Xception65 | 5.398 | 4.215 | 8.611 | 28.702 | -| Xception65
_deeplab | 5.317 | 3.688 | 6.168 | 23.108 | -| Xception71 | 2.732 | 5.033 | 8.948 | 33.857 | -| InceptionV4 | 6.172 | 7.558 | 9.527 | 24.021 | - ## FP32预测速度 -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| GoogLeNet | 1.436 | 2.904 | 3.800 | 9.049 | -| Xception41 | 3.402 | 7.889 | 14.953 | 56.142 | -| Xception41
_deeplab | 3.778 | 8.396 | 15.449 | 58.735 | -| Xception65 | 6.802 | 13.935 | 34.301 | 87.256 | -| Xception65
_deeplab | 8.583 | 12.132 | 22.917 | 87.983 | -| Xception71 | 6.156 | 14.359 | 27.360 | 107.282 | -| InceptionV4 | 10.384 | 17.438 | 23.312 | 68.777 | +| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +|------------------------|-----------|-------------------|--------------------------| +| GoogLeNet | 224 | 256 | 1.807 | +| Xception41 | 299 | 320 | 3.972 | +| Xception41
_deeplab | 299 | 320 | 4.408 | +| Xception65 | 299 | 320 | 6.174 | +| Xception65
_deeplab | 299 | 320 | 6.464 | +| Xception71 | 299 | 320 | 6.782 | +| InceptionV4 | 299 | 320 | 11.141 | diff --git a/docs/zh_CN/models/Mobile.md b/docs/zh_CN/models/Mobile.md index 3331e02dd2077d89fc19669fe3652ae927fa0a72..1a6406f16d8c1e0edd05620318c277bcfe3aef2b 100644 --- a/docs/zh_CN/models/Mobile.md +++ b/docs/zh_CN/models/Mobile.md @@ -10,11 +10,10 @@ ShuffleNet系列网络是旷视提出的轻量化网络结构，到目前为止 MobileNetV3是Google于2019年提出的一种基于NAS的新的轻量级网络，为了进一步提升效果，将relu和sigmoid激活函数分别替换为hard_swish与hard_sigmoid激活函数，同时引入了一些专门减小网络计算量的改进策略。 ![](../../images/models/mobile_arm_top1.png) ![](../../images/models/mobile_arm_storage.png) -![](../../images/models/mobile_trt.png) +![](../../images/models/mobile_trt.png.flops.png) +![](../../images/models/mobile_trt.png.params.png) 目前PaddleClas开源的的移动端系列的预训练模型一共有32个，其指标如图所示。从图片可以看出，越新的轻量级模型往往有更优的表现，MobileNetV3代表了目前最新的轻量级神经网络结构。在MobileNetV3中，作者为了获得更高的精度，在global-avg-pooling后使用了1x1的卷积。该操作大幅提升了参数量但对计算量影响不大，所以如果从存储角度评价模型的优异程度，MobileNetV3优势不是很大，但由于其更小的计算量，使得其有更快的推理速度。此外，我们模型库中的ssld蒸馏模型表现优异，从各个考量角度下，都刷新了当前轻量级模型的精度。由于MobileNetV3模型结构复杂，分支较多，对GPU并不友好，GPU预测速度不如MobileNetV1。 -**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 @@ -54,78 +53,42 @@ MobileNetV3是Google于2019年提出的一种基于NAS的新的轻量级网络 | ShuffleNetV2_swish | 0.700 | 0.892 | | | 0.290 | 2.260 | -## FP16预测速度 - -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| MobileNetV1_x0_25 | 0.236 | 0.258 | 0.281 | 0.556 | -| MobileNetV1_x0_5 | 0.246 | 0.318 | 0.364 | 0.845 | -| MobileNetV1_x0_75 | 0.303 | 0.380 | 0.512 | 1.164 | -| MobileNetV1 | 0.340 | 0.426 | 0.601 | 1.578 | -| MobileNetV1_ssld | 0.340 | 0.426 | 0.601 | 1.578 | -| MobileNetV2_x0_25 | 0.432 | 0.488 | 0.532 | 0.967 | -| MobileNetV2_x0_5 | 0.475 | 0.564 | 0.654 | 1.296 | -| MobileNetV2_x0_75 | 0.553 | 0.653 | 0.821 | 1.761 | -| MobileNetV2 | 0.610 | 0.738 | 0.931 | 2.115 | -| MobileNetV2_x1_5 | 0.731 | 0.966 | 1.252 | 3.152 | -| MobileNetV2_x2_0 | 0.870 | 1.123 | 1.494 | 3.910 | -| MobileNetV2_ssld | 0.610 | 0.738 | 0.931 | 2.115 | -| MobileNetV3_large_
x1_25 | 2.004 | 2.223 | 2.433 | 5.954 | -| MobileNetV3_large_
x1_0 | 1.943 | 2.203 | 2.113 | 4.823 | -| MobileNetV3_large_
x0_75 | 2.107 | 2.266 | 2.120 | 3.968 | -| MobileNetV3_large_
x0_5 | 1.942 | 2.178 | 2.179 | 2.936 | -| MobileNetV3_large_
x0_35 | 1.994 | 2.407 | 2.285 | 2.420 | -| MobileNetV3_small_
x1_25 | 1.876 | 2.141 | 2.118 | 3.423 | -| MobileNetV3_small_
x1_0 | 1.751 | 2.160 | 2.203 | 2.830 | -| MobileNetV3_small_
x0_75 | 1.856 | 2.235 | 2.166 | 2.464 | -| MobileNetV3_small_
x0_5 | 1.773 | 2.304 | 2.242 | 2.133 | -| MobileNetV3_small_
x0_35 | 1.870 | 2.392 | 2.323 | 2.101 | -| MobileNetV3_large_
x1_0_ssld | 1.943 | 2.203 | 2.113 | 4.823 | | -| MobileNetV3_small_
x1_0_ssld | 1.751 | 2.160 | 2.203 | 2.830 | -| ShuffleNetV2 | 1.134 | 1.068 | 1.199 | 2.558 | -| ShuffleNetV2_x0_25 | 0.911 | 0.953 | 0.948 | 1.327 | -| ShuffleNetV2_x0_33 | 0.853 | 1.072 | 0.958 | 1.398 | -| ShuffleNetV2_x0_5 | 0.858 | 1.059 | 1.084 | 1.620 | -| ShuffleNetV2_x1_5 | 1.040 | 1.153 | 1.394 | 3.452 | -| ShuffleNetV2_x2_0 | 1.061 | 1.316 | 1.694 | 4.485 | -| ShuffleNetV2_swish | 1.688 | 1.958 | 1.707 | 3.711 | - ## FP32预测速度 -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| MobileNetV1_x0_25 | 0.233 | 0.372 | 0.424 | 0.930 | -| MobileNetV1_x0_5 | 0.281 | 0.532 | 0.677 | 1.808 | -| MobileNetV1_x0_75 | 0.344 | 0.733 | 0.960 | 2.920 | -| MobileNetV1 | 0.420 | 0.963 | 1.462 | 4.769 | -| MobileNetV1_ssld | 0.420 | 0.963 | 1.462 | 4.769 | -| MobileNetV2_x0_25 | 0.718 | 0.738 | 0.775 | 1.482 | -| MobileNetV2_x0_5 | 0.818 | 0.975 | 1.107 | 2.481 | -| MobileNetV2_x0_75 | 0.830 | 1.104 | 1.514 | 3.629 | -| MobileNetV2 | 0.889 | 1.346 | 1.875 | 4.711 | -| MobileNetV2_x1_5 | 1.221 | 1.982 | 2.951 | 7.645 | -| MobileNetV2_x2_0 | 1.546 | 2.625 | 3.734 | 10.429 | -| MobileNetV2_ssld | 0.889 | 1.346 | 1.875 | 4.711 | -| MobileNetV3_large_
x1_25 | 2.113 | 2.377 | 3.114 | 7.332 | -| MobileNetV3_large_
x1_0 | 1.991 | 2.380 | 2.517 | 5.826 | -| MobileNetV3_large_
x0_75 | 2.105 | 2.454 | 2.336 | 4.611 | -| MobileNetV3_large_
x0_5 | 1.978 | 2.603 | 2.291 | 3.306 | -| MobileNetV3_large_
x0_35 | 2.017 | 2.469 | 2.316 | 2.558 | -| MobileNetV3_small_
x1_25 | 1.915 | 2.411 | 2.295 | 3.742 | -| MobileNetV3_small_
x1_0 | 1.915 | 2.889 | 2.862 | 3.022 | -| MobileNetV3_small_
x0_75 | 1.941 | 2.358 | 2.232 | 2.602 | -| MobileNetV3_small_
x0_5 | 1.872 | 2.364 | 2.238 | 2.061 | -| MobileNetV3_small_
x0_35 | 1.889 | 2.407 | 2.328 | 2.127 | -| MobileNetV3_large_
x1_0_ssld | 1.991 | 2.380 | 2.517 | 5.826 | -| MobileNetV3_small_
x1_0_ssld | 1.915 | 2.889 | 2.862 | 3.022 | -| ShuffleNetV2 | 1.328 | 1.211 | 1.440 | 3.210 | -| ShuffleNetV2_x0_25 | 0.905 | 0.908 | 0.924 | 1.284 | -| ShuffleNetV2_x0_33 | 0.871 | 1.073 | 0.891 | 1.416 | -| ShuffleNetV2_x0_5 | 0.852 | 1.150 | 1.093 | 1.702 | -| ShuffleNetV2_x1_5 | 0.874 | 1.470 | 1.889 | 4.490 | -| ShuffleNetV2_x2_0 | 1.443 | 1.908 | 2.556 | 6.864 | -| ShuffleNetV2_swish | 1.694 | 1.856 | 2.101 | 3.942 | +| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +|--------------------------------------|-----------|-------------------|--------------------------| +| MobileNetV1_x0_25 | 224 | 256 | 0.492 | +| MobileNetV1_x0_5 | 224 | 256 | 0.599 | +| MobileNetV1_x0_75 | 224 | 256 | 0.695 | +| MobileNetV1 | 224 | 256 | 0.739 | +| MobileNetV1_ssld | 224 | 256 | 0.739 | +| MobileNetV2_x0_25 | 224 | 256 | 1.014 | +| MobileNetV2_x0_5 | 224 | 256 | 1.216 | +| MobileNetV2_x0_75 | 224 | 256 | 1.392 | +| MobileNetV2 | 224 | 256 | 1.153 | +| MobileNetV2_x1_5 | 224 | 256 | 1.516 | +| MobileNetV2_x2_0 | 224 | 256 | 1.819 | +| MobileNetV2_ssld | 224 | 256 | 1.153 | +| MobileNetV3_large_
x1_25 | 224 | 256 | 3.070 | +| MobileNetV3_large_
x1_0 | 224 | 256 | 3.173 | +| MobileNetV3_large_
x0_75 | 224 | 256 | 2.928 | +| MobileNetV3_large_
x0_5 | 224 | 256 | 2.979 | +| MobileNetV3_large_
x0_35 | 224 | 256 | 2.987 | +| MobileNetV3_small_
x1_25 | 224 | 256 | 3.003 | +| MobileNetV3_small_
x1_0 | 224 | 256 | 3.168 | +| MobileNetV3_small_
x0_75 | 224 | 256 | 2.974 | +| MobileNetV3_small_
x0_5 | 224 | 256 | 2.199 | +| MobileNetV3_small_
x0_35 | 224 | 256 | 2.240 | +| MobileNetV3_large_
x1_0_ssld | 224 | 256 | 3.173 | +| MobileNetV3_small_
x1_0_ssld | 224 | 256 | 3.168 | +| ShuffleNetV2 | 224 | 256 | 1.861 | +| ShuffleNetV2_x0_25 | 224 | 256 | 1.410 | +| ShuffleNetV2_x0_33 | 224 | 256 | 1.271 | +| ShuffleNetV2_x0_5 | 224 | 256 | 1.389 | +| ShuffleNetV2_x1_5 | 224 | 256 | 1.239 | +| ShuffleNetV2_x2_0 | 224 | 256 | 2.152 | +| ShuffleNetV2_swish | 224 | 256 | 2.150 | ## CPU预测速度和存储大小 diff --git a/docs/zh_CN/models/Others.md b/docs/zh_CN/models/Others.md index e57a6c9752e8e10560865abe328cf639fc0df769..1b5d2c11d881e89f0158692563519fe0f663e8b8 100644 --- a/docs/zh_CN/models/Others.md +++ b/docs/zh_CN/models/Others.md @@ -3,8 +3,6 @@ ## 概述 正在持续更新中...... -DarkNet53在预测时，图像的crop_size设置为256，resize_short_size设置为256；其余模型在预测时，图像的crop_size设置为224，resize_short_size设置为256。 - ## 精度、FLOPS和参数量 @@ -22,31 +20,18 @@ DarkNet53在预测时，图像的crop_size设置为256，resize_short_size设置 | ResNet50_ACNet
_deploy | 0.767 | 0.932 | | | 8.190 | 25.550 | -## FP16预测速度 - -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| AlexNet | 0.684 | 0.740 | 0.810 | 1.481 | -| SqueezeNet1_0 | 0.545 | 0.841 | 1.146 | 3.501 | -| SqueezeNet1_1 | 0.473 | 0.575 | 0.805 | 1.862 | -| VGG11 | 1.096 | 1.655 | 2.396 | 6.728 | -| VGG13 | 1.216 | 2.059 | 3.056 | 9.468 | -| VGG16 | 1.518 | 2.594 | 4.019 | 12.145 | -| VGG19 | 1.817 | 3.124 | 4.886 | 14.958 | -| DarkNet53 | 2.150 | 2.627 | 3.422 | 10.092 | | -| ResNet50_ACNet
_deploy | 2.748 | 3.178 | 3.823 | 8.369 | - ## FP32预测速度 -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| AlexNet | 0.682 | 0.875 | 1.196 | 3.196 | -| SqueezeNet1_0 | 0.530 | 1.072 | 1.652 | 5.338 | -| SqueezeNet1_1 | 0.439 | 0.787 | 1.164 | 2.973 | -| VGG11 | 1.575 | 3.638 | 6.427 | 23.227 | -| VGG13 | 1.859 | 4.832 | 8.832 | 32.946 | -| VGG16 | 2.316 | 6.420 | 11.936 | 44.719 | -| VGG19 | 2.775 | 8.013 | 14.925 | 57.272 | -| DarkNet53 | 2.648 | 5.727 | 9.616 | 33.664 | | -| ResNet50_ACNet
_deploy | 4.544 | 6.873 | 9.627 | 28.283 | + +| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +|---------------------------|-----------|-------------------|----------------------| +| AlexNet | 224 | 256 | 1.176 | +| SqueezeNet1_0 | 224 | 256 | 0.860 | +| SqueezeNet1_1 | 224 | 256 | 0.763 | +| VGG11 | 224 | 256 | 1.867 | +| VGG13 | 224 | 256 | 2.148 | +| VGG16 | 224 | 256 | 2.616 | +| VGG19 | 224 | 256 | 3.076 | +| DarkNet53 | 256 | 256 | 3.139 | +| ResNet50_ACNet
_deploy | 224 | 256 | 5.626 | diff --git a/docs/zh_CN/models/ResNet_and_vd.md b/docs/zh_CN/models/ResNet_and_vd.md index 7fc5e816235f9322fb5963b890aaefa39e3dec49..89a55d4ff4d7272dbedefb9eb571b3ffc24d970b 100644 --- a/docs/zh_CN/models/ResNet_and_vd.md +++ b/docs/zh_CN/models/ResNet_and_vd.md @@ -9,7 +9,16 @@ ResNet系列模型是在2015年提出的，一举在ILSVRC2015比赛中取得冠 本次发布ResNet系列的模型包括ResNet50，ResNet50_vd，ResNet50_vd_ssld，ResNet200_vd等14个预训练模型。在训练层面上，ResNet的模型采用了训练ImageNet的标准训练流程，而其余改进版模型采用了更多的训练策略，如learning rate的下降方式采用了cosine decay，引入了label smoothing的标签正则方式，在数据预处理加入了mixup的操作，迭代总轮数从120个epoch增加到200个epoch。 其中，ResNet50_vd_v2与ResNet50_vd_ssld采用了知识蒸馏，保证模型结构不变的情况下，进一步提升了模型的精度，具体地，ResNet50_vd_v2的teacher模型是ResNet152_vd（top1准确率80.59%），数据选用的是ImageNet-1k的训练集，ResNet50_vd_ssld的teacher模型是ResNeXt101_32x16d_wsl（top1准确率84.2%），数据选用结合了ImageNet-1k的训练集和ImageNet-22k挖掘的400万数据。知识蒸馏的具体方法正在持续更新中。 -![](../../images/models/ResNet.png) + + +该系列模型的FLOPS、参数量以及fp32预测耗时如下图所示。 + +![](../../images/models/ResNet.png.flops.png) + +![](../../images/models/ResNet.png.params.png) + +![](../../images/models/ResNet.png.fp32.png) + 通过上述曲线可以看出，层数越多，准确率越高，但是相应的参数量、计算量和延时都会增加。ResNet50_vd_ssld通过用更强的teacher和更多的数据，将其在ImageNet-1k上的验证集top-1精度进一步提高，达到了82.39%，刷新了ResNet50系列模型的精度。 **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。 @@ -32,43 +41,27 @@ ResNet系列模型是在2015年提出的，一举在ILSVRC2015比赛中取得冠 | ResNet152_vd | 0.806 | 0.953 | | | 23.530 | 60.210 | | ResNet200_vd | 0.809 | 0.953 | | | 30.530 | 74.740 | | ResNet50_vd_ssld | 0.824 | 0.961 | | | 8.670 | 25.580 | +| ResNet101_vd_ssld | 0.835 | 0.968 | | | 16.100 | 44.570 | -## FP16预测速度 - -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| ResNet18 | 0.966 | 1.076 | 1.263 | 2.656 | -| ResNet18_vd | 1.002 | 1.163 | 1.392 | 3.045 | -| ResNet34 | 1.798 | 1.959 | 2.269 | 4.716 | -| ResNet34_vd | 1.839 | 2.011 | 2.482 | 4.767 | -| ResNet50 | 1.892 | 2.146 | 2.692 | 6.411 | -| ResNet50_vc | 1.903 | 2.094 | 2.677 | 6.096 | -| ResNet50_vd | 1.918 | 2.273 | 2.833 | 6.978 | -| ResNet50_vd_v2 | 1.918 | 2.273 | 2.833 | 6.978 | -| ResNet101 | 3.790 | 4.128 | 4.789 | 10.913 | -| ResNet101_vd | 3.853 | 4.229 | 5.001 | 11.437 | -| ResNet152 | 5.523 | 5.871 | 6.710 | 15.258 | -| ResNet152_vd | 5.503 | 6.003 | 7.001 | 15.716 | -| ResNet200_vd | 7.270 | 7.595 | 8.802 | 19.516 | -| ResNet50_vd_ssld | 1.918 | 2.273 | 2.833 | 6.978 | ## FP32预测速度 -| Models | batch_size=1
(ms) | batch_size=4
(ms) | batch_size=8
(ms) | batch_size=32
(ms) | -|:--:|:--:|:--:|:--:|:--:| -| ResNet18 | 1.127 | 1.428 | 2.352 | 7.780 | -| ResNet18_vd | 1.142 | 1.532 | 2.584 | 8.441 | -| ResNet34 | 1.936 | 2.409 | 4.197 | 14.442 | -| ResNet34_vd | 1.948 | 2.526 | 4.403 | 15.133 | -| ResNet50 | 2.630 | 4.393 | 6.491 | 20.449 | -| ResNet50_vc | 2.728 | 4.413 | 6.618 | 21.183 | -| ResNet50_vd | 2.649 | 4.522 | 6.771 | 21.552 | -| ResNet50_vd_v2 | 2.649 | 4.522 | 6.771 | 21.552 | -| ResNet101 | 4.747 | 8.015 | 11.555 | 36.739 | -| ResNet101_vd | 4.735 | 8.111 | 11.820 | 37.155 | -| ResNet152 | 6.618 | 11.471 | 16.580 | 51.792 | -| ResNet152_vd | 6.626 | 11.613 | 16.843 | 53.645 | -| ResNet200_vd | 8.540 | 14.770 | 21.554 | 69.053 | -| ResNet50_vd_ssld | 2.649 | 4.522 | 6.771 | 21.552 | +| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +|------------------|-----------|-------------------|--------------------------| +| ResNet18 | 224 | 256 | 1.499 | +| ResNet18_vd | 224 | 256 | 1.603 | +| ResNet34 | 224 | 256 | 2.272 | +| ResNet34_vd | 224 | 256 | 2.343 | +| ResNet50 | 224 | 256 | 2.939 | +| ResNet50_vc | 224 | 256 | 3.041 | +| ResNet50_vd | 224 | 256 | 3.165 | +| ResNet50_vd_v2 | 224 | 256 | 3.165 | +| ResNet101 | 224 | 256 | 5.314 | +| ResNet101_vd | 224 | 256 | 5.252 | +| ResNet152 | 224 | 256 | 7.205 | +| ResNet152_vd | 224 | 256 | 7.200 | +| ResNet200_vd | 224 | 256 | 8.885 | +| ResNet50_vd_ssld | 224 | 256 | 3.165 | +| ResNet101_vd_ssld | 224 | 256 | 5.252 | diff --git a/docs/zh_CN/models/SEResNext_and_Res2Net.md b/docs/zh_CN/models/SEResNext_and_Res2Net.md index 577b1b3473c9e8d4f16b0e8d40fc583e838bba02..946315548583985f28bf8dab497ef0a3b9fc06c8 100644 --- a/docs/zh_CN/models/SEResNext_and_Res2Net.md +++ b/docs/zh_CN/models/SEResNext_and_Res2Net.md @@ -6,9 +6,18 @@ ResNeXt是ResNet的典型变种网络之一，ResNeXt发表于2017年的CVPR会 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。 -![](../../images/models/SeResNeXt.png) + +该系列模型的FLOPS、参数量以及fp32预测耗时如下图所示。 + +![](../../images/models/SeResNeXt.png.flops.png) + +![](../../images/models/SeResNeXt.png.params.png) + +![](../../images/models/SeResNeXt.png.fp32.png) + 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。 @@ -42,61 +51,32 @@ Res2Net是2019年提出的一种全新的对ResNet的改进方案，该方案可 | 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 | +| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +|-----------------------|-----------|-------------------|--------------------------| +| Res2Net50_26w_4s | 224 | 256 | 4.148 | +| Res2Net50_vd_26w_4s | 224 | 256 | 4.172 | +| Res2Net50_14w_8s | 224 | 256 | 5.113 | +| Res2Net101_vd_26w_4s | 224 | 256 | 7.327 | +| Res2Net200_vd_26w_4s | 224 | 256 | 12.806 | +| ResNeXt50_32x4d | 224 | 256 | 10.964 | +| ResNeXt50_vd_32x4d | 224 | 256 | 7.566 | +| ResNeXt50_64x4d | 224 | 256 | 13.905 | +| ResNeXt50_vd_64x4d | 224 | 256 | 14.321 | +| ResNeXt101_32x4d | 224 | 256 | 14.915 | +| ResNeXt101_vd_32x4d | 224 | 256 | 14.885 | +| ResNeXt101_64x4d | 224 | 256 | 28.716 | +| ResNeXt101_vd_64x4d | 224 | 256 | 28.398 | +| ResNeXt152_32x4d | 224 | 256 | 22.996 | +| ResNeXt152_vd_32x4d | 224 | 256 | 22.729 | +| ResNeXt152_64x4d | 224 | 256 | 46.705 | +| ResNeXt152_vd_64x4d | 224 | 256 | 46.395 | +| SE_ResNet18_vd | 224 | 256 | 1.694 | +| SE_ResNet34_vd | 224 | 256 | 2.786 | +| SE_ResNet50_vd | 224 | 256 | 3.749 | +| SE_ResNeXt50_32x4d | 224 | 256 | 8.924 | +| SE_ResNeXt50_vd_32x4d | 224 | 256 | 9.011 | +| SE_ResNeXt101_32x4d | 224 | 256 | 19.204 | +| SENet154_vd | 224 | 256 | 50.406 | diff --git a/docs/zh_CN/models/models_intro.md b/docs/zh_CN/models/models_intro.md index f0240a18819f3f78613ee2cca86a906510e7fe05..e8e08fc1027dd5377cfbaaed51fd7bc7ac808cd6 100644 --- a/docs/zh_CN/models/models_intro.md +++ b/docs/zh_CN/models/models_intro.md @@ -2,7 +2,26 @@ ## 概述 -基于ImageNet1k分类数据集，PaddleClas支持的23种系列分类网络结构以及对应的117个图像分类预训练模型如下所示，训练技巧、每个系列网络结构的简单介绍和性能评估将在相应章节展现。GPU评估环境基于V100和TensorRT，CPU的评估环境基于骁龙855（SD855）。 +基于ImageNet1k分类数据集，PaddleClas支持的23种系列分类网络结构以及对应的117个图像分类预训练模型如下所示，训练技巧、每个系列网络结构的简单介绍和性能评估将在相应章节展现。 + +## 评估环境 +* CPU的评估环境基于骁龙855（SD855）。 +* GPU评估环境基于V100和TensorRT，评估脚本如下。 + +```shell +#!/usr/bin/env bash + +export PYTHONPATH=$PWD:$PYTHONPATH + +python tools/infer/predict.py \ + --model_file='pretrained/infer/model' \ + --params_file='pretrained/infer/params' \ + --enable_benchmark=True \ + --model_name=ResNet50_vd \ + --use_tensorrt=True \ + --use_fp16=False \ + --batch_size=1 +``` ![](../../images/models/main_fps_top1.png) ![](../../images/models/mobile_arm_top1.png) @@ -25,6 +44,7 @@ - [ResNet152_vd](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet152_vd_pretrained.tar) - [ResNet200_vd](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet200_vd_pretrained.tar) - [ResNet50_vd_ssld](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_ssld_pretrained.tar) + - [ResNet101_vd_ssld](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_ssld_pretrained.tar) - 移动端系列 diff --git a/tools/infer/predict.py b/tools/infer/predict.py index 4ef4af5c1032defecac15073c1bcdc96f0a22f24..c6bfabafc430c45b290eb84cfcbe514537b1ad6b 100644 --- a/tools/infer/predict.py +++ b/tools/infer/predict.py @@ -109,11 +109,6 @@ def main(): operators = create_operators() predictor = create_predictor(args) - inputs = preprocess(args.image_file, operators) - inputs = np.expand_dims( - inputs, axis=0).repeat( - args.batch_size, axis=0).copy() - input_names = predictor.get_input_names() input_tensor = predictor.get_input_tensor(input_names[0])