当前位置: 首页 > 工具软件 > model-zoo > 使用案例 >

MMDetection 基准测试 和 Model Zoo | 三

饶骁
2023-12-01

作者|open-mmlab 编译|Flin 来源|Github

基准测试 和 Model Zoo

环境

硬件
  • 8 个 NVIDIA Tesla V100 GPUs
  • Intel Xeon 4114 CPU @ 2.20GHz
软件环境
  • Python 3.6 / 3.7
  • PyTorch 1.1
  • CUDA 9.0.176
  • CUDNN 7.0.4
  • NCCL 2.1.15

镜像站点

我们使用AWS作为托管model zoo的主要站点,并在阿里云上维护镜像。 你可以在模型网址中把https://s3.ap-northeast-2.amazonaws.com/open-mmlab替换为https://open-mmlab.oss-cn-beijing.aliyuncs.com。

常用设置

  • 所有FPN基准和RPN-C4基准均使用8个GPU进行训练,批处理大小为16(每个GPU 2张图像)。其他C4基线使用8个批处理大小为8的GP​​U进行了训练(每个GPU 1张图像)。
  • 所有模型都在coco_2017_train上训练以及在coco_2017_val测试。
  • 我们使用分布式训练,并且BN层统计信息是固定的。
  • 我们采用与Detectron相同的训练时间表。1x表示12个epoch,而2x表示24个epoch,这比Detectron的迭代次数略少,并且可以忽略不计。
  • ImageNet上所有pytorch样式的预训练主干都来自PyTorchmodel zoo。
  • 为了与其他代码库进行公平比较,我们将GPU内存报告 torch.cuda.max_memory_allocated()为所有8个GPU 的最大值。请注意,此值通常小于nvidia-smi显示的值。
  • 我们将推理时间报告为总体时间,包括数据加载,网络转发和后处理。

基线

具有不同主干的更多模型将添加到model zoo。

RPN
BackboneStyleLr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)AR1000Download
R-50-C4caffe1x--20.551.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_r50_caffe_c4_1x-ea7d3428.pth)
R-50-C4caffe2x2.20.1720.352.2model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_r50_caffe_c4_2x-c6d5b958.pth)
R-50-C4pytorch1x--20.150.2model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_r50_c4_1x-eb38972b.pth)
R-50-C4pytorch2x--20.051.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_r50_c4_2x-3d4c1e14.pth)
R-50-FPNcaffe1x3.30.25316.958.2-
R-50-FPNpytorch1x3.50.27617.757.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_r50_fpn_1x_20181010-4a9c0712.pth)
R-50-FPNpytorch2x---57.6model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_r50_fpn_2x_20181010-88a4a471.pth)
R-101-FPNcaffe1x5.20.37913.959.4-
R-101-FPNpytorch1x5.40.39614.458.6model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_r101_fpn_1x_20181129-f50da4bd.pth)
R-101-FPNpytorch2x---59.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_r101_fpn_2x_20181129-e42c6c9a.pth)
X-101-32x4d-FPNpytorch1x6.60.58911.859.4model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_x101_32x4d_fpn_1x_20181218-7e379d26.pth)
X-101-32x4d-FPNpytorch2x---59.9model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_x101_32x4d_fpn_2x_20181218-0510af40.pth)
X-101-64x4d-FPNpytorch1x9.50.9558.359.8model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_x101_64x4d_fpn_1x_20181218-c1a24f1f.pth)
X-101-64x4d-FPNpytorch2x---60.0model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/rpn_x101_64x4d_fpn_2x_20181218-c22bdd70.pth)
Faster R-CNN
BackboneStyleLr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)box APDownload
R-50-C4caffe1x--9.534.9model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_r50_caffe_c4_1x-75ecfdfa.pth)
R-50-C4caffe2x4.00.399.336.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_r50_caffe_c4_2x-71c67f27.pth)
R-50-C4pytorch1x--9.333.9model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_r50_c4_1x-642cf91f.pth)
R-50-C4pytorch2x--9.435.9model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_r50_c4_2x-6e4fdf4f.pth)
R-50-FPNcaffe1x3.60.33313.536.6-
R-50-FPNpytorch1x3.80.35313.636.4model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_r50_fpn_1x_20181010-3d1b3351.pth)
R-50-FPNpytorch2x---37.7model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_r50_fpn_2x_20181010-443129e1.pth)
R-101-FPNcaffe1x5.50.46511.538.8-
R-101-FPNpytorch1x5.70.47411.938.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_r101_fpn_1x_20181129-d1468807.pth)
R-101-FPNpytorch2x---39.4model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_r101_fpn_2x_20181129-73e7ade7.pth)
X-101-32x4d-FPNpytorch1x6.90.67210.340.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_x101_32x4d_fpn_1x_20181218-ad81c133.pth)
X-101-32x4d-FPNpytorch2x---40.4model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_x101_32x4d_fpn_2x_20181218-0ed58946.pth)
X-101-64x4d-FPNpytorch1x9.81.0407.341.3model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_x101_64x4d_fpn_1x_20181218-c9c69c8f.pth)
X-101-64x4d-FPNpytorch2x---40.7model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/faster_rcnn_x101_64x4d_fpn_2x_20181218-fe94f9b8.pth)
HRNetV2p-W18pytorch1x---36.1model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/faster_rcnn_hrnetv2p_w18_1x_20190522-e368c387.pth)
HRNetV2p-W18pytorch2x---38.3model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/faster_rcnn_hrnetv2p_w18_2x_20190810-9c8615d5.pth)
HRNetV2p-W32pytorch1x---39.5model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/faster_rcnn_hrnetv2p_w32_1x_20190522-d22f1fef.pth)
HRNetV2p-W32pytorch2x---40.6model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/faster_rcnn_hrnetv2p_w32_2x_20190810-24e8912a.pth)
HRNetV2p-W48pytorch1x---40.9model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/faster_rcnn_hrnetv2p_w48_1x_20190820-5c6d0903.pth)
HRNetV2p-W48pytorch2x---41.5model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/faster_rcnn_hrnetv2p_w48_2x_20190820-79fb8bfc.pth)
Mask R-CNN
BackboneStyleLr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)box APmask APDownload
R-50-C4caffe1x--8.135.931.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_r50_caffe_c4_1x-02a4ad3b.pth)
R-50-C4caffe2x4.20.438.137.932.9model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_r50_caffe_c4_2x-d150973a.pth)
R-50-C4pytorch1x--7.935.131.2model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_r50_c4_1x-a83bdd40.pth)
R-50-C4pytorch2x--8.037.232.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_r50_c4_2x-3cf169a9.pth)
R-50-FPNcaffe1x3.80.43010.237.434.3-
R-50-FPNpytorch1x3.90.45310.637.334.2model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_r50_fpn_1x_20181010-069fa190.pth)
R-50-FPNpytorch2x---38.535.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_r50_fpn_2x_20181010-41d35c05.pth)
R-101-FPNcaffe1x5.70.5349.439.936.1-
R-101-FPNpytorch1x5.80.5719.539.435.9model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_r101_fpn_1x_20181129-34ad1961.pth)
R-101-FPNpytorch2x---40.336.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_r101_fpn_2x_20181129-a254bdfc.pth)
X-101-32x4d-FPNpytorch1x7.10.7598.341.137.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_x101_32x4d_fpn_1x_20181218-44e635cc.pth)
X-101-32x4d-FPNpytorch2x---41.437.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_x101_32x4d_fpn_2x_20181218-f023dffa.pth)
X-101-64x4d-FPNpytorch1x10.01.1026.542.138.0model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_x101_64x4d_fpn_1x_20181218-cb159987.pth)
X-101-64x4d-FPNpytorch2x---42.037.7model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/mask_rcnn_x101_64x4d_fpn_2x_20181218-ea936e44.pth)
HRNetV2p-W18pytorch1x---37.334.2model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/mask_rcnn_hrnetv2p_w18_1x_20190522-c8ad459f.pth)
HRNetV2p-W18pytorch2x---39.235.7model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/mask_rcnn_hrnetv2p_w18_2x_20190810-1e4747eb.pth)
HRNetV2p-W32pytorch1x---40.736.8model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/mask_rcnn_hrnetv2p_w32_1x_20190522-374aaa00.pth)
HRNetV2p-W32pytorch2x---41.737.5model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/mask_rcnn_hrnetv2p_w32_2x_20190810-773eca75.pth)
HRNetV2p-W48pytorch1x---42.438.1model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/mask_rcnn_hrnetv2p_w48_1x_20190820-0923d1ad.pth)
HRNetV2p-W48pytorch2x---42.938.3model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/mask_rcnn_hrnetv2p_w48_2x_20190820-70df51b2.pth)
Fast R-CNN (有预先计算的proposals)
BackboneStyle类型Lr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)box APmask APDownload
R-50-C4caffeFaster1x--6.735.0-model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_rcnn_r50_caffe_c4_1x-0ef9a60b.pth)
R-50-C4caffeFaster2x3.80.346.636.4-model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_rcnn_r50_c4_2x-657a9fc6.pth)
R-50-C4pytorchFaster1x--6.334.2-model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_rcnn_r50_c4_1x-2bc00ca9.pth)
R-50-C4pytorchFaster2x--6.135.8-model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_rcnn_r50_caffe_c4_2x-9171d0fc.pth)
R-50-FPNcaffeFaster1x3.30.24218.436.6--
R-50-FPNpytorchFaster1x3.50.25016.535.8-model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_rcnn_r50_fpn_1x_20181010-08160859.pth)
R-50-C4caffeMask1x--8.135.931.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_mask_rcnn_r50_caffe_c4_1x-b43f7f3c.pth)
R-50-C4caffeMask2x4.20.438.137.932.9model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_mask_rcnn_r50_caffe_c4_2x-e3580184.pth)
R-50-C4pytorchMask1x--7.935.131.2model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_mask_rcnn_r50_c4_1x-bc7fa8c8.pth)
R-50-C4pytorchMask2x--8.037.232.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_mask_rcnn_r50_fpn_2x_20181010-5048cb03.pth)
R-50-FPNpytorchFaster2x---37.1-model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_rcnn_r50_fpn_2x_20181010-d263ada5.pth)
R-101-FPNcaffeFaster1x5.20.35514.438.6--
R-101-FPNpytorchFaster1x5.40.38813.238.1-model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_rcnn_r101_fpn_1x_20181129-ffaa2eb0.pth)
R-101-FPNpytorchFaster2x---38.8-model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_rcnn_r101_fpn_2x_20181129-9dba92ce.pth)
R-50-FPNcaffeMask1x3.40.32812.837.334.5-
R-50-FPNpytorchMask1x3.50.34612.736.834.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_mask_rcnn_r50_fpn_1x_20181010-e030a38f.pth)
R-50-FPNpytorchMask2x---37.934.8model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_mask_rcnn_r50_fpn_2x_20181010-5048cb03.pth)
R-101-FPNcaffeMask1x5.20.42911.239.436.1-
R-101-FPNpytorchMask1x5.40.46210.938.935.8model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_mask_rcnn_r101_fpn_1x_20181129-2273fa9b.pth)
R-101-FPNpytorchMask2x---39.936.4model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/fast_mask_rcnn_r101_fpn_2x_20181129-bf63ec5e.pth)
RetinaNet
BackboneStyleLr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)box APDownload
R-50-FPNcaffe1x3.40.28512.535.8-
R-50-FPNpytorch1x3.60.30812.135.6model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/retinanet_r50_fpn_1x_20181125-7b0c2548.pth)
R-50-FPNpytorch2x---36.4model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/retinanet_r50_fpn_2x_20190616-75574209.pth)
R-101-FPNcaffe1x5.30.41010.437.8-
R-101-FPNpytorch1x5.50.42910.937.7model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/retinanet_r101_fpn_1x_20181129-f016f384.pth)
R-101-FPNpytorch2x---38.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/retinanet_r101_fpn_2x_20181129-72c14526.pth)
X-101-32x4d-FPNpytorch1x6.70.6329.339.0model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/retinanet_x101_32x4d_fpn_1x_20190501-967812ba.pth)
X-101-32x4d-FPNpytorch2x---39.3model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/retinanet_x101_32x4d_fpn_2x_20181218-8596452d.pth)
X-101-64x4d-FPNpytorch1x9.60.9937.040.0model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/retinanet_x101_64x4d_fpn_1x_20181218-a0a22662.pth)
X-101-64x4d-FPNpytorch2x---39.6model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/retinanet_x101_64x4d_fpn_2x_20181218-5e88d045.pth)
Cascade R-CNN
BackboneStyleLr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)box APDownload
R-50-C4caffe1x8.70.925.038.7model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_r50_caffe_c4_1x-7c85c62b.pth)
R-50-FPNcaffe1x3.90.46410.940.5-
R-50-FPNpytorch1x4.10.45511.940.4model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_r50_fpn_1x_20190501-3b6211ab.pth)
R-50-FPNpytorch20e---41.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_r50_fpn_20e_20181123-db483a09.pth)
R-101-FPNcaffe1x5.80.5699.642.4-
R-101-FPNpytorch1x6.00.58410.342.0model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_r101_fpn_1x_20181129-d64ebac7.pth)
R-101-FPNpytorch20e---42.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_r101_fpn_20e_20181129-b46dcede.pth)
X-101-32x4d-FPNpytorch1x7.20.7708.943.6model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_x101_32x4d_fpn_1x_20190501-af628be5.pth)
X-101-32x4d-FPNpytorch20e---44.0model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_x101_32x4d_fpn_2x_20181218-28f73c4c.pth)
X-101-64x4d-FPNpytorch1x10.01.1336.744.5model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_x101_64x4d_fpn_1x_20181218-e2dc376a.pth)
X-101-64x4d-FPNpytorch20e---44.7model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_rcnn_x101_64x4d_fpn_2x_20181218-5add321e.pth)
HRNetV2p-W18pytorch20e---41.2model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/cascade_rcnn_hrnetv2p_w18_20e_20190810-132012d0.pth)
HRNetV2p-W32pytorch20e---43.7model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/cascade_rcnn_hrnetv2p_w32_20e_20190522-55bec4ee.pth)
HRNetV2p-W48pytorch20e---44.6model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/cascade_rcnn_hrnetv2p_w48_20e_20190810-f40ed8e1.pth)
Cascade Mask R-CNN
BackboneStyleLr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)box APmask APDownload
R-50-C4caffe1x9.10.994.539.332.8model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_r50_caffe_c4_1x-f72cc254.pth)
R-50-FPNcaffe1x5.10.6927.640.935.5-
R-50-FPNpytorch1x5.30.6837.441.235.7model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_r50_fpn_1x_20181123-88b170c9.pth)
R-50-FPNpytorch20e---42.336.6model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_r50_fpn_20e_20181123-6e0c9713.pth)
R-101-FPNcaffe1x7.00.8037.243.137.2-
R-101-FPNpytorch1x7.20.8076.842.637.0model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_r101_fpn_1x_20181129-64f00602.pth)
R-101-FPNpytorch20e---43.337.6model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_r101_fpn_20e_20181129-cb85151d.pth)
X-101-32x4d-FPNpytorch1x8.40.9766.644.438.2model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_x101_32x4d_fpn_1x_20181218-1d944c89.pth)
X-101-32x4d-FPNpytorch20e---44.738.6model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_x101_32x4d_fpn_20e_20181218-761a3473.pth)
X-101-64x4d-FPNpytorch1x11.41.335.345.439.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_x101_64x4d_fpn_1x_20190501-827e0a70.pth)
X-101-64x4d-FPNpytorch20e---45.739.4model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/cascade_mask_rcnn_x101_64x4d_fpn_20e_20181218-630773a7.pth)
HRNetV2p-W18pytorch20e---41.936.4model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/cascade_mask_rcnn_hrnetv2p_w18_20e_20190810-054fb7bf.pth)
HRNetV2p-W32pytorch20e---44.538.5model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/cascade_mask_rcnn_hrnetv2p_w32_20e_20190810-76f61cd0.pth)
HRNetV2p-W48pytorch20e---46.039.5model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/cascade_mask_rcnn_hrnetv2p_w48_20e_20190810-d04a1415.pth)

注意s:

  • 20e级联(掩码)R-CNN中的时间表指示在第16和19个epoch减少lr,总共减少20个epoch。
混合任务级联(HTC)
BackboneStyleLr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)box APmask APDownload
R-50-FPNpytorch1x7.40.9364.142.137.3model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/htc/htc_r50_fpn_1x_20190408-878c1712.pth)
R-50-FPNpytorch20e---43.238.1model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/htc/htc_r50_fpn_20e_20190408-c03b7015.pth)
R-101-FPNpytorch20e9.31.0514.044.939.4model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/htc/htc_r101_fpn_20e_20190408-a2e586db.pth)
X-101-32x4d-FPNpytorch20e5.80.7693.846.140.3model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/htc/htc_x101_32x4d_fpn_20e_20190408-9eae4d0b.pth)
X-101-64x4d-FPNpytorch20e7.51.1203.546.940.8model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/htc/htc_x101_64x4d_fpn_20e_20190408-497f2561.pth)
HRNetV2p-W18pytorch20e---43.137.9model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/htc_hrnetv2p_w18_20e_20190810-d70072af.pth)
HRNetV2p-W32pytorch20e---45.339.6model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/htc_hrnetv2p_w32_20e_20190810-82f9ef5a.pth)
HRNetV2p-W48pytorch20e---46.840.7model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/htc_hrnetv2p_w48_20e_20190810-f6d2c3fd.pth)
HRNetV2p-W48pytorch28e---47.041.0model(https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/hrnet/htc_hrnetv2p_w48_28e_20190810-a4274b38.pth)

注意:

SSD
BackboneSizeStyleLr schd内存 (GB)训练时间 (s/iter)最短时间 (fps)box APDownload
VGG16300caffe120e3.50.25625.9 / 34.625.7model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/ssd300_coco_vgg16_caffe_120e_20181221-84d7110b.pth)
VGG16512caffe120e7.60.41220.7 / 25.429.3model(https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/ssd512_coco_vgg16_caffe_120e_20181221-d48b0be8.pth)

注意:

  • cudnn.benchmark设置为True用于SSD训练和测试。
  • 对于batch size= 1和batch size= 8,报告推理时间。
  • 由于模型参数和nms,COCO和VOC的速度有所不同。
组规范化(GN)

有关详细信息,请参考组规范化(https://github.com/open-mmlab/mmdetection/blob/master/configs/gn)。

权重标准化

有关详细信息,请参考权重标准化(https://github.com/open-mmlab/mmdetection/blob/master/configs/gn+ws)。

可变形卷积v2

有关详细信息,请参阅可变形卷积网络(https://github.com/open-mmlab/mmdetection/blob/master/configs/dcn)。

CARAFE:功能的内容感知重组

有关详细信息,请参考CARAFE(https://github.com/open-mmlab/mmdetection/blob/master/configs/carafe)。

Instaboost

有关详细信息,请参考Instaboost(https://github.com/open-mmlab/mmdetection/blob/master/configs/instaboost)。

Libra R-CNN

有关详细信息,请参考Libra R-CNN(https://github.com/open-mmlab/mmdetection/blob/master/configs/libra_rcnn)。

Guided Anchoring

有关详细信息,请参阅Guided Anchoring(https://github.com/open-mmlab/mmdetection/blob/master/configs/guided_anchoring)。

FCOS

有关详细信息,请参阅FCOS(https://github.com/open-mmlab/mmdetection/blob/master/configs/fcos)。

FoveaBox

有关详细信息,请参考FoveaBox(https://github.com/open-mmlab/mmdetection/blob/master/configs/foveabox)。

RepPoints

有关详细信息,请参考RepPoints(https://github.com/open-mmlab/mmdetection/blob/master/configs/reppoints)。

FreeAnchor

有关详细信息,请参考FreeAnchor(https://github.com/open-mmlab/mmdetection/blob/master/configs/free_anchor)。

Grid R-CNN (plus)

有关详细信息,请参考Grid R-CNN(https://github.com/open-mmlab/mmdetection/blob/master/configs/grid_rcnn)。

GHM

有关详细信息,请参阅GHM(https://github.com/open-mmlab/mmdetection/blob/master/configs/ghm)。

GCNet

有关详细信息,请参考GCNet(https://github.com/open-mmlab/mmdetection/blob/master/configs/gcnet)。

HRNet

有关详细信息,请参考HRNet(https://github.com/open-mmlab/mmdetection/blob/master/configs/hrnet)。

Mask Scoring R-CNN

有关详细信息,请参考Mask Scoring R-CNN(https://github.com/open-mmlab/mmdetection/blob/master/configs/ms_rcnn)。

Train from Scratch

有关详细信息,请参考 重新思考ImageNet预训练(https://github.com/open-mmlab/mmdetection/blob/master/configs/scratch)。

NAS-FPN

有关详细信息,请参阅NAS-FPN(https://github.com/open-mmlab/mmdetection/blob/master/configs/nas_fpn)。

ATSS

有关详细信息,请参考ATSS(https://github.com/open-mmlab/mmdetection/blob/master/configs/atss)。

其他数据集

我们还对PASCAL VOC(https://github.com/open-mmlab/mmdetection/blob/master/configs/pascal_voc),Cityscapes(https://github.com/open-mmlab/ mmdetection / blob / master / configs / cityscapes)和WIDER FACE(https://github.com/open-mmlab/mmdetection/blob/master/configs/wider_face)的一些方法进行了基准测试。

与 Detectron 和 maskrcnn-benchmark 的比较

我们将mmdetection与Detectron(https://github.com/facebookresearch/Detectron) 和maskrcnn-benchmark(https://github.com/facebookresearch/maskrcnn-benchmark)进行比较。使用的主干是R-50-FPN。

通常来说,mmdetection与Detectron相比具有3个优势。

  • 更高的性能(尤其是在mask AP方面)
  • 更快的训练速度
  • 高效记忆
性能

Detectron和maskrcnn-benchmark使用Caffe风格的ResNet作为主干。我们使用caffe样式(权重从(https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md#imagenet-pretrained-models) 和pytorch样式(权重来自官方model zoo)ResNet主干报告结果,表示为pytorch样式结果 / caffe样式结果。

我们发现,pytorch风格的ResNet通常比caffe风格的ResNet收敛慢,因此在1倍进度中导致结果略低,但2倍进度的最终结果则较高。

类型Lr schdDetectronmaskrcnn-benchmarkmmdetection
RPN1x57.2-57.1 / 58.2
2x--57.6 / -
Faster R-CNN1x36.736.836.4 / 36.6
2x37.9-37.7 / -
Mask R-CNN1x37.7 & 33.937.8 & 34.237.3 & 34.2 / 37.4 & 34.3
2x38.6 & 34.5-38.5 & 35.1 / -
Fast R-CNN1x36.4-35.8 / 36.6
2x36.8-37.1 / -
Fast R-CNN (w/mask)1x37.3 & 33.7-36.8 & 34.1 / 37.3 & 34.5
2x37.7 & 34.0-37.9 & 34.8 / -
训练速度

训练速度以s/iter为单位。越低越好。

类型Detectron (P1001)maskrcnn-benchmark (V100)mmdetection (V1002)
RPN0.416-0.253
Faster R-CNN0.5440.3530.333
Mask R-CNN0.8890.4540.430
Fast R-CNN0.285-0.242
Fast R-CNN (w/mask)0.377-0.328
  • 1。Facebook的Big Basin服务器(P100 / V100)比我们使用的服务器稍快。mmdetection在FB的服务器上也可以稍快一些地运行。

  • 2。为了公平比较,我们在此处列出了caffe的结果。

推理速度

推理速度在单个GPU上以fps(img / s)进行测量。越高越好。

类型Detectron (P100)maskrcnn-benchmark (V100)mmdetection (V100)
RPN12.5-16.9
Faster R-CNN10.37.913.5
Mask R-CNN8.57.710.2
Fast R-CNN12.5-18.4
Fast R-CNN (w/mask)9.9-12.8
训练内存
类型Detectronmaskrcnn-benchmarkmmdetection
RPN6.4-3.3
Faster R-CNN7.24.43.6
Mask R-CNN8.65.23.8
Fast R-CNN6.0-3.3
Fast R-CNN (w/mask)7.9-3.4

毫无疑问,maskrcnn基准测试和mmdetection比Detectron的存储效率更高,而主要优点是PyTorch本身。我们还执行一些内存优化来推动它向前发展。

请注意,Caffe2和PyTorch具有不同的API,以通过不同的实现获取内存使用情况。对于所有代码库,nvidia-smi显示的内存使用量均大于上表中报告的数字。

原文链接:https://mmdetection.readthedocs.io/en/latest/MODEL_ZOO.html

欢迎关注磐创AI博客站: http://panchuang.net/

sklearn机器学习中文官方文档: http://sklearn123.com/

欢迎关注磐创博客资源汇总站: http://docs.panchuang.net/

 类似资料: