From a764cd1eeaf27212e563a975ef7183d319cfc25d Mon Sep 17 00:00:00 2001 From: Kaipeng Deng Date: Mon, 9 Mar 2020 12:30:36 +0800 Subject: [PATCH] [cherry-pick] polish slim model zoo (#313) * polish slim README --- slim/README.md | 218 +++++++++++++++++++++++++++++++------------------ 1 file changed, 138 insertions(+), 80 deletions(-) diff --git a/slim/README.md b/slim/README.md index bc63703ef..9161c7b0d 100644 --- a/slim/README.md +++ b/slim/README.md @@ -43,27 +43,41 @@ ### YOLOv3 on COCO -| 骨架网络 | 剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 | Box AP | 下载 | -| :----------------| :-------: | :---------: | :------------: | :------: | :-----: | :-----------------------------------------------------: | -| ResNet50-vd-dcn | sensity | 18.41% | 15.46% | 608 | 39.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune1x.tar) | -| ResNet50-vd-dcn | r578 | 43.69% | 36.61% | 608 | 38.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune578.tar) | -| MobileNetV1 | sensity | 28.76% | 28.54% | 608 | 30.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | -| MobileNetV1 | sensity | 28.76% | 28.54% | 416 | 29.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | -| MobileNetV1 | sensity | 28.76% | 28.54% | 320 | 27.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | -| MobileNetV1 | r578 | 67.56% | 66.90% | 608 | 27.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) | -| MobileNetV1 | r578 | 67.56% | 66.90% | 416 | 26.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) | -| MobileNetV1 | r578 | 67.56% | 66.90% | 320 | 24.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) | +| 骨架网络 | 剪裁策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | Box AP | 下载 | +| :----------------| :-------: | :------------: | :-------------: | :------: | :--------: | :-----------------------------------------------------: | +| ResNet50-vd-dcn | baseline | 44.71 | 176.82 | 608 | 39.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn.tar) | +| ResNet50-vd-dcn | sensity | 37.53(-16.06%) | 149.49(-15.46%) | 608 | 39.8(+0.7) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune1x.tar) | +| ResNet50-vd-dcn | r578 | 29.98(-32.94%) | 112.08(-36.61%) | 608 | 38.3(-0.8) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune578.tar) | +| MobileNetV1 | baseline | 20.64 | 94.60 | 608 | 29.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | baseline | 9.66 | 94.60 | 416 | 29.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | baseline | 5.72 | 94.60 | 320 | 27.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | sensity | 13.57(-34.27%) | 67.60(-28.54%) | 608 | 30.2(+0.9) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | +| MobileNetV1 | sensity | 6.35(-34.27%) | 67.60(-28.54%) | 416 | 29.7(+0.4) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | +| MobileNetV1 | sensity | 3.76(-34.27%) | 67.60(-28.54%) | 320 | 27.2(+0.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | +| MobileNetV1 | r578 | 6.27(-69.64%) | 31.30(-66.90%) | 608 | 27.8(-1.5) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) | +| MobileNetV1 | r578 | 2.93(-69.64%) | 31.30(-66.90%) | 416 | 26.8(-2.5) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) | +| MobileNetV1 | r578 | 1.74(-69.64%) | 31.30(-66.90%) | 320 | 24.0(-3.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) | + +- YOLO v3在训练阶段对minibatch采用随机reshape，可以采用相同的模型权重不同尺寸图片，表中`YOLOv3-MobileNetV1`提供了在`608/416/320`三种不同尺寸下的精度结果 +- 在使用`sensity`剪裁策略下，`YOLOv3-ResNet50-vd-dcn`和`YOLOv3-MobileNetV1`分别减少了`16.06%`和`34.27%`的FLOPs，输入图像尺寸为608时精度分别提高`0.7`和`0.9` +- 在使用`r578`剪裁策略下，`YOLOv3-ResNet50-vd-dcn`和`YOLOv3-MobileNetV1`分别减少了`32.98%`和`69.64%`的FLOPs，输入图像尺寸为608时精度分别降低`0.8`和`1.5` ### YOLOv3 on Pascal VOC -| 骨架网络 | 剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 | Box AP | 下载 | -| :----------------| :-------: | :---------: | :------------: | :------: | :-----: | :-----------------------------------------------------: | -| MobileNetV1 | sensity | 34.55% | 28.75% | 608 | 78.4 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | -| MobileNetV1 | sensity | 34.55% | 28.75% | 416 | 78.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | -| MobileNetV1 | sensity | 34.55% | 28.75% | 320 | 76.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | -| MobileNetV1 | r578 | 69.57% | 67.00% | 608 | 77.6 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | -| MobileNetV1 | r578 | 69.57% | 67.00% | 416 | 77.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | -| MobileNetV1 | r578 | 69.57% | 67.00% | 320 | 75.5 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | +| 骨架网络 | 剪裁策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | Box AP | 下载 | +| :----------------| :-------: | :------------: | :-------------: | :------: | :--------: | :-----------------------------------------------------: | +| MobileNetV1 | baseline | 20.20 | 93.37 | 608 | 76.2 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | baseline | 9.46 | 93.37 | 416 | 76.7 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | baseline | 5.60 | 93.37 | 320 | 75.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | sensity | 13.22(-34.55%) | 66.53(-28.74%) | 608 | 78.4(+2.2) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | +| MobileNetV1 | sensity | 6.19(-34.55%) | 66.53(-28.74%) | 416 | 78.7(+2.0) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | +| MobileNetV1 | sensity | 3.66(-34.55%) | 66.53(-28.74%) | 320 | 76.1(+0.8) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | +| MobileNetV1 | r578 | 6.15(-69.57%) | 30.81(-67.00%) | 608 | 77.6(+1.4) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | +| MobileNetV1 | r578 | 2.88(-69.57%) | 30.81(-67.00%) | 416 | 77.7(+1.0) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | +| MobileNetV1 | r578 | 1.70(-69.57%) | 30.81(-67.00%) | 320 | 75.5(+0.2) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | + +- YOLO v3在训练阶段对minibatch采用随机reshape，可以采用相同的模型权重不同尺寸图片，表中`YOLOv3-MobileNetV1`提供了在`608/416/320`三种不同尺寸下的精度结果 +- 在使用`sensity`和`r578`剪裁策略下，`YOLOv3-MobileNetV1`分别减少了`34.55%`和`69.57%`的FLOPs，输入图像尺寸为608时精度分别提高`2.2`和`1.4` ### 蒸馏通道剪裁模型 @@ -71,20 +85,34 @@ COCO数据集上蒸馏通道剪裁模型库如下。 -| 骨架网络 | 剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 | teacher模型 | Box AP | 下载 | -| :----------------| :-------: | :---------: | :------------: | :------: | :--------------------- | :-----: | :-----------------------------------------------------: | -| ResNet50-vd-dcn | r578 | 43.69% | 36.61% | 608 | YOLOv3-ResNet50-vd-dcn | 39.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune578_distill.tar) | -| MobileNetV1 | r578 | 67.56% | 66.90% | 608 | YOLOv3-ResNet34 | 29.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | -| MobileNetV1 | r578 | 67.56% | 66.90% | 416 | YOLOv3-ResNet34 | 28.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | -| MobileNetV1 | r578 | 67.56% | 66.90% | 320 | YOLOv3-ResNet34 | 25.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | +| 骨架网络 | 剪裁策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | teacher模型 | Box AP | 下载 | +| :----------------| :-------: | :------------: | :-------------: | :------: | :--------------------------: | :--------: | :-----------------------------------------------------: | +| ResNet50-vd-dcn | baseline | 44.71 | 176.82 | 608 | - | 39.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn.tar) | +| ResNet50-vd-dcn | r578 | 29.98(-32.94%) | 112.08(-36.61%) | 608 | YOLOv3-ResNet50-vd-dcn(39.1) | 39.7(+0.6) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune578_distill.tar) | +| MobileNetV1 | baseline | 20.64 | 94.60 | 608 | - | 29.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | baseline | 9.66 | 94.60 | 416 | - | 29.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | baseline | 5.72 | 94.60 | 320 | - | 27.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | r578 | 6.27(-69.64%) | 31.30(-66.90%) | 608 | YOLOv3-ResNet34(36.2) | 29.0(-0.3) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | +| MobileNetV1 | r578 | 2.93(-69.64%) | 31.30(-66.90%) | 416 | YOLOv3-ResNet34(34.3) | 28.0(-1.3) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | +| MobileNetV1 | r578 | 1.74(-69.64%) | 31.30(-66.90%) | 320 | YOLOv3-ResNet34(31.4) | 25.1(-2.0) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | + +- YOLO v3在训练阶段对minibatch采用随机reshape，可以采用相同的模型权重不同尺寸图片，表中`YOLOv3-MobileNetV1`提供了在`608/416/320`三种不同尺寸下的精度结果 +- 在使用`r578`剪裁策略并使用`YOLOv3-ResNet50-vd-dcn`作为teacher模型蒸馏，`YOLOv3-ResNet50-vd-dcn`模型减少了`32.94%`的FLOPs，输入图像尺寸为608时精度提高`0.6` +- 在使用`r578`剪裁策略并使用`YOLOv3-ResNet34`作为teacher模型蒸馏下，`YOLOv3-MobileNetV1`模型减少了`69.64%`的FLOPs，输入图像尺寸为608时精度降低`0.3` Pascal VOC数据集上蒸馏通道剪裁模型库如下。 -| 骨架网络 | 剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 | teacher模型 | Box AP | 下载 | -| :----------------| :-------: | :---------: | :------------: | :------: | :--------------------- | :-----: | :-----------------------------------------------------: | -| MobileNetV1 | r578 | 69.57% | 67.00% | 608 | YOLOv3-ResNet34 | 78.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | -| MobileNetV1 | r578 | 69.57% | 67.00% | 416 | YOLOv3-ResNet34 | 78.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | -| MobileNetV1 | r578 | 69.57% | 67.00% | 320 | YOLOv3-ResNet34 | 76.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | +| 骨架网络 | 剪裁策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | teacher模型 | Box AP | 下载 | +| :----------------| :-------: | :------------: | :-------------: | :------: | :--------------------: | :--------: | :-----------------------------------------------------: | +| MobileNetV1 | baseline | 20.20 | 93.37 | 608 | - | 76.2 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | baseline | 9.46 | 93.37 | 416 | - | 76.7 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | baseline | 5.60 | 93.37 | 320 | - | 75.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | r578 | 6.15(-69.57%) | 30.81(-67.00%) | 608 | YOLOv3-ResNet34(82.6) | 78.8(+2.6) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | +| MobileNetV1 | r578 | 2.88(-69.57%) | 30.81(-67.00%) | 416 | YOLOv3-ResNet34(81.9) | 78.7(+2.0) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | +| MobileNetV1 | r578 | 1.70(-69.57%) | 30.81(-67.00%) | 320 | YOLOv3-ResNet34(80.1) | 76.3(+2.0) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | + +- YOLO v3在训练阶段对minibatch采用随机reshape，可以采用相同的模型权重不同尺寸图片，表中`YOLOv3-MobileNetV1`提供了在`608/416/320`三种不同尺寸下的精度结果 +- 在使用`r578`剪裁策略并使用`YOLOv3-ResNet34`作为teacher模型蒸馏下，`YOLOv3-MobileNetV1`模型减少了`69.57%`的FLOPs，输入图像尺寸为608时精度提高`2.6` ### YOLOv3通道剪裁模型推理时延 @@ -92,24 +120,20 @@ Pascal VOC数据集上蒸馏通道剪裁模型库如下。 - Tesla P4时延为单卡并开启TensorRT推理时延 - 高通835/高通855/麒麟970时延为使用PaddleLite部署，使用`arm8`架构并使用4线程(4 Threads)推理时延 -| 骨架网络 | 数据集 | 剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 | Tesla P4 | 麒麟970 | 高通835 | 高通855 | -| :--------------- | :----: | :------: | :---------: | :------------: | :------: | :------: | :-----: | :-----: | :-----: | -| MobileNetV1 | VOC | baseline | - | - | 608 | 16.556 | 748.404 | 734.970 | 289.878 | -| MobileNetV1 | VOC | baseline | - | - | 416 | 9.031 | 371.214 | 349.065 | 140.877 | -| MobileNetV1 | VOC | baseline | - | - | 320 | 6.235 | 221.705 | 200.498 | 80.515 | -| MobileNetV1 | VOC | r578 | 69.57% | 67.00% | 608 | 10.064 | 314.531 | 323.537 | 123.414 | -| MobileNetV1 | VOC | r578 | 69.57% | 67.00% | 416 | 5.478 | 151.562 | 146.014 | 56.420 | -| MobileNetV1 | VOC | r578 | 69.57% | 67.00% | 320 | 3.880 | 91.132 | 87.440 | 31.470 | -| ResNet50-vd-dcn | COCO | baseline | - | - | 608 | 36.127 | - | - | - | -| ResNet50-vd-dcn | COCO | baseline | - | - | 416 | 20.437 | - | - | - | -| ResNet50-vd-dcn | COCO | baseline | - | - | 320 | 14.037 | - | - | - | -| ResNet50-vd-dcn | COCO | sensity | 18.41% | 15.46% | 608 | 33.245 | - | - | - | -| ResNet50-vd-dcn | COCO | sensity | 18.41% | 15.46% | 416 | 19.246 | - | - | - | -| ResNet50-vd-dcn | COCO | sensity | 18.41% | 15.46% | 320 | 13.656 | - | - | - | -| ResNet50-vd-dcn | COCO | r578 | 43.69% | 36.61% | 608 | 29.138 | - | - | - | -| ResNet50-vd-dcn | COCO | r578 | 43.69% | 36.61% | 416 | 16.439 | - | - | - | -| ResNet50-vd-dcn | COCO | r578 | 43.69% | 36.61% | 320 | 11.339 | - | - | - | - +| 骨架网络 | 数据集 | 剪裁策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | Tesla P4 | 麒麟970 | 高通835 | 高通855 | +| :--------------- | :----: | :------: | :------------: | :-------------: | :------: | :-------------: | :--------------: | :--------------: | :--------------: | +| MobileNetV1 | VOC | baseline | 20.20 | 93.37 | 608 | 16.556 | 748.404 | 734.970 | 289.878 | +| MobileNetV1 | VOC | baseline | 9.46 | 93.37 | 416 | 9.031 | 371.214 | 349.065 | 140.877 | +| MobileNetV1 | VOC | baseline | 5.60 | 93.37 | 320 | 6.235 | 221.705 | 200.498 | 80.515 | +| MobileNetV1 | VOC | r578 | 6.15(-69.57%) | 30.81(-67.00%) | 608 | 10.064(-39.21%) | 314.531(-57.97%) | 323.537(-55.98%) | 123.414(-57.43%) | +| MobileNetV1 | VOC | r578 | 2.88(-69.57%) | 30.81(-67.00%) | 416 | 5.478(-39.34%) | 151.562(-59.17%) | 146.014(-58.17%) | 56.420(-59.95%) | +| MobileNetV1 | VOC | r578 | 1.70(-69.57%) | 30.81(-67.00%) | 320 | 3.880(-37.77%) | 91.132(-58.90%) | 87.440(-56.39%) | 31.470(-60.91%) | +| ResNet50-vd-dcn | COCO | baseline | 44.71 | 176.82 | 608 | 36.127 | - | - | - | +| ResNet50-vd-dcn | COCO | sensity | 37.53(-16.06%) | 149.49(-15.46%) | 608 | 33.245(-7.98%) | - | - | - | +| ResNet50-vd-dcn | COCO | r578 | 29.98(-32.94%) | 112.08(-36.61%) | 608 | 29.138(-19.35%) | - | - | - | + +- 在使用`r578`剪裁策略下，`YOLOv3-MobileNetV1`模型减少了`69.57%`的FLOPs，输入图像尺寸为608时在单卡Tesla P4(TensorRT)推理时间减少`39.21%`，在麒麟970/高通835/高通855上推理时延分别减少`57.97%`, `55.98%`和`57.43%` +- 在使用`sensity`和`r578`剪裁策略下，`YOLOv3-ResNet50-vd-dcn`模型分别减少了`16.06%`和`32.94%`的FLOPs，输入图像尺寸为608时在单卡Tesla P4(TensorRT)推理时间分别减少`7.98%`和`19.35%` ## 蒸馏模型库 @@ -117,22 +141,36 @@ Pascal VOC数据集上蒸馏通道剪裁模型库如下。 - 蒸馏模型训练时teacher模型使用[PaddleDetection模型库](https://paddledetection.readthedocs.io/zh/latest/MODEL_ZOO_cn.html)发布的模型权重作为预训练权重。 - 蒸馏模型训练时student模型使用backbone的预训练权重 +- 蒸馏策略`l2_distiil`为使用teacher模型和student模型特征图的L2损失作为蒸馏损失进行蒸馏，为`slim/distillation/distill.py`的默认策略 +- 蒸馏策略`split_distiil`为使用YOLOv3细粒度损失进行蒸馏，通过`-o use_fine_grained_loss=true`指定 ### YOLOv3 on COCO -| 骨架网络 | 蒸馏策略 | 输入尺寸 | Box AP | 下载 | -| :----------------| :-----------: | :------: |:------: | :-----------------------------------------------------: | -| MobileNetV1 | split_distiil | 608 | 31.4 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | -| MobileNetV1 | split_distiil | 416 | 30.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | -| MobileNetV1 | split_distiil | 320 | 27.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | +| 骨架网络 | 蒸馏策略 | 输入尺寸 | teacher模型 | Box AP | 下载 | +| :----------------| :-----------: | :------: | :--------------------: | :----------: | :-----------------------------------------------------: | +| MobileNetV1 | baseline | 608 | - | 29.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | baseline | 416 | - | 29.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | baseline | 320 | - | 27.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | split_distiil | 608 | YOLOv3-ResNet34(36.2) | 31.4(+2.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | +| MobileNetV1 | split_distiil | 416 | YOLOv3-ResNet34(34.3) | 30.0(+0.7) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | +| MobileNetV1 | split_distiil | 320 | YOLOv3-ResNet34(31.4) | 27.1(+0.0) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | + +- YOLO v3在训练阶段对minibatch采用随机reshape，可以采用相同的模型权重不同尺寸图片，表中`YOLOv3-MobileNetV1`提供了在`608/416/320`三种不同尺寸下的精度结果 +- 在使用`YOLOv3-ResNet34`模型通过`split_distiil`策略蒸馏下，输入图像尺寸为608时`YOLOv3-MobileNetV1`模型精度提高`2.1` ### YOLOv3 on Pascal VOC -| 骨架网络 | 蒸馏策略 | 输入尺寸 | Box AP | 下载 | -| :----------------| :-----------: | :------: |:------: | :-----------------------------------------------------: | -| MobileNetV1 | l2_distiil | 608 | 79.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) | -| MobileNetV1 | l2_distiil | 416 | 78.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) | -| MobileNetV1 | l2_distiil | 320 | 75.5 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) | +| 骨架网络 | 蒸馏策略 | 输入尺寸 | teacher模型 | Box AP | 下载 | +| :----------------| :-----------: | :------: | :--------------------: | :--------: | :-----------------------------------------------------: | +| MobileNetV1 | baseline | 608 | - | 76.2 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | baseline | 416 | - | 76.7 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | baseline | 320 | - | 75.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNetV1 | l2_distiil | 608 | YOLOv3-ResNet34(82.6) | 79.0(+2.8) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) | +| MobileNetV1 | l2_distiil | 416 | YOLOv3-ResNet34(81.9) | 78.2(+1.5) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) | +| MobileNetV1 | l2_distiil | 320 | YOLOv3-ResNet34(80.1) | 75.5(+0.2) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) | + +- YOLO v3在训练阶段对minibatch采用随机reshape，可以采用相同的模型权重不同尺寸图片，表中`YOLOv3-MobileNetV1`提供了在`608/416/320`三种不同尺寸下的精度结果 +- 在使用`YOLOv3-ResNet34`模型通过`l2_distiil`策略蒸馏下，输入图像尺寸为608时`YOLOv3-MobileNetV1`模型精度提高`2.8` ## 量化模型库 @@ -142,29 +180,49 @@ Pascal VOC数据集上蒸馏通道剪裁模型库如下。 ### YOLOv3 on COCO -| 骨架网络 | 预训练权重 | 量化策略 | 输入尺寸 | Box AP | 下载 | -| :----------------| :--------: | :------: | :------: |:------: | :-----------------------------------------------------: | -| MobileNetV1 | ImageNet | post | 608 | 27.9 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | -| MobileNetV1 | ImageNet | post | 416 | 28.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | -| MobileNetV1 | ImageNet | post | 320 | 26.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | -| MobileNetV1 | ImageNet | aware | 608 | 28.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | -| MobileNetV1 | ImageNet | aware | 416 | 28.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | -| MobileNetV1 | ImageNet | aware | 320 | 25.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | -| ResNet34 | ImageNet | post | 608 | 35.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) | -| ResNet34 | ImageNet | aware | 608 | 35.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | -| ResNet34 | ImageNet | aware | 416 | 33.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | -| ResNet34 | ImageNet | aware | 320 | 30.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | -| R50vd-dcn | object365 | aware | 608 | 40.6 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | -| R50vd-dcn | object365 | aware | 416 | 37.5 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | -| R50vd-dcn | object365 | aware | 320 | 34.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | +| 骨架网络 | 预训练权重 | 量化策略 | 输入尺寸 | Box AP | 下载 | +| :----------------| :--------: | :------: | :------: | :--------: | :-----------------------------------------------------: | +| MobileNetV1 | ImageNet | baseline | 608 | 29.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | ImageNet | baseline | 416 | 29.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | ImageNet | baseline | 320 | 27.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNetV1 | ImageNet | post | 608 | 27.9(-1.4) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | +| MobileNetV1 | ImageNet | post | 416 | 28.0(-1.3) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | +| MobileNetV1 | ImageNet | post | 320 | 26.0(-1.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | +| MobileNetV1 | ImageNet | aware | 608 | 28.1(-1.2) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | +| MobileNetV1 | ImageNet | aware | 416 | 28.2(-1.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | +| MobileNetV1 | ImageNet | aware | 320 | 25.8(-1.3) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | +| ResNet34 | ImageNet | baseline | 608 | 36.2 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) | +| ResNet34 | ImageNet | baseline | 416 | 34.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) | +| ResNet34 | ImageNet | baseline | 320 | 31.4 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) | +| ResNet34 | ImageNet | post | 608 | 35.7(-0.5) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) | +| ResNet34 | ImageNet | aware | 608 | 35.2(-1.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | +| ResNet34 | ImageNet | aware | 416 | 33.3(-1.0) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | +| ResNet34 | ImageNet | aware | 320 | 30.3(-1.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | +| R50vd-dcn | object365 | baseline | 608 | 41.4 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) | +| R50vd-dcn | object365 | aware | 608 | 40.6(-0.8) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | +| R50vd-dcn | object365 | aware | 416 | 37.5(-) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | +| R50vd-dcn | object365 | aware | 320 | 34.1(-) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | + +- YOLO v3在训练阶段对minibatch采用随机reshape，可以采用相同的模型权重不同尺寸图片，表中部分模型提供了在`608/416/320`三种不同尺寸下的精度结果 +- `YOLOv3-MobileNetV1`使用离线(post)和在线(aware)两种量化方式，输入图像尺寸为608时精度分别降低`1.4`和`1.2` +- `YOLOv3-ResNet34`使用离线(post)和在线(aware)两种量化方式，输入图像尺寸为608时精度分别降低`0.5`和`1.1` +- `YOLOv3-R50vd-dcn`使用在线(aware)量化方式，输入图像尺寸为608时精度降低`0.8` ### BlazeFace on WIDER FACE -| 模型 | 量化策略 | 输入尺寸 | Easy Set | Medium Set | Hard Set | 下载 | -| :--------------- | :------: | :------: | :------: | :--------: | :------: | :-----------------------------------------------------: | -| BlazeFace | post | 640 | 87.8 | 85.1 | 74.9 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_post.tar) | -| BlazeFace | aware | 640 | 90.5 | 87.9 | 77.6 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_aware.tar) | -| BlazeFace-Lite | post | 640 | 89.4 | 86.7 | 75.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_post.tar) | -| BlazeFace-Lite | aware | 640 | 89.7 | 87.3 | 77.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_aware.tar) | -| BlazeFace-NAS | post | 640 | 81.6 | 78.3 | 63.6 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_post.tar) | -| BlazeFace-NAS | aware | 640 | 83.1 | 79.7 | 64.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_aware.tar) | +| 模型 | 量化策略 | 输入尺寸 | Easy Set | Medium Set | Hard Set | 下载 | +| :--------------- | :------: | :------: | :--------: | :--------: | :--------: | :-----------------------------------------------------: | +| BlazeFace | baseline | 640 | 91.5 | 89.2 | 79.7 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) | +| BlazeFace | post | 640 | 87.8(-3.7) | 85.1(-3.9) | 74.9(-4.8) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_post.tar) | +| BlazeFace | aware | 640 | 90.5(-1.0) | 87.9(-1.3) | 77.6(-2.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_aware.tar) | +| BlazeFace-Lite | baseline | 640 | 90.9 | 88.5 | 78.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_lite.tar) | +| BlazeFace-Lite | post | 640 | 89.4(-1.5) | 86.7(-1.8) | 75.7(-2.4) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_post.tar) | +| BlazeFace-Lite | aware | 640 | 89.7(-1.2) | 87.3(-1.2) | 77.0(-1.1) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_aware.tar) | +| BlazeFace-NAS | baseline | 640 | 83.7 | 80.7 | 65.8 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) | +| BlazeFace-NAS | post | 640 | 81.6(-2.1) | 78.3(-2.4) | 63.6(-2.2) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_post.tar) | +| BlazeFace-NAS | aware | 640 | 83.1(-0.6) | 79.7(-1.0) | 64.2(-1.6) | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_aware.tar) | + +- `BlazeFace`系列模型中在线(aware)量化性能明显优于离线(post)量化 +- `BlazeFace`模型使用在线(aware)量化方式，在`Easy/Medium/Hard`数据集上精度分别降低`1.0`, `1.3`和`2.1` +- `BlazeFace-Lite`模型使用在线(aware)量化方式，在`Easy/Medium/Hard`数据集上精度分别降低`1.2`, `1.2`和`1.1` +- `BlazeFace-NAS`模型使用在线(aware)量化方式，在`Easy/Medium/Hard`数据集上精度分别降低`0.6`, `1.0`和`1.6` -- GitLab