mirrors
/
mmdeploy
mirror of https://github.com/open-mmlab/mmdeploy.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
mmdeploy/docs/benchmark.md

28 KiB
Raw Blame History

Benchmark

Backends

CPU: ncnn, ONNXRuntime GPU: TensorRT, PPLNN

Platform

  • Ubuntu 18.04
  • Cuda 11.3
  • TensorRT 7.2.3.4
  • Docker 20.10.8
  • NVIDIA tesla T4 tensor core GPU for TensorRT.

Other settings

  • Static graph
  • Batch size 1
  • Synchronize devices after each inference.
  • We count the average inference performance of 100 images of the dataset.
  • Warm up. For classification, we warm up 1010 iters. For other codebases, we warm up 10 iters.
  • Input resolution varies for different datasets of different codebases. All inputs are real images except for mmediting because the dataset is not large enough.

Latency benchmark

Users can directly test the speed through how_to_measure_performance_of_models.md. And here is the benchmark in our environment.

MMCls with 1x3x224x224 input
TensorRT PPLNN
Model Input fp32 fp16 in8 fp16 model config file
latency (ms) FPS latency (ms) FPS latency (ms) FPS latency (ms) FPS
ResNet 1x3x224x224 2.97 336.90 1.26 791.89 1.21 829.66 1.30 768.28 $MMCLS_DIR/configs/resnet/resnet50_b32x8_imagenet.py
ResNeXt 1x3x224x224 4.31 231.93 1.42 703.42 1.37 727.42 1.36 737.67 $MMCLS_DIR/configs/resnext/resnext50_32x4d_b32x8_imagenet.py
SE-ResNet 1x3x224x224 3.41 293.64 1.66 600.73 1.51 662.90 1.91 524.07 $MMCLS_DIR/configs/seresnet/seresnet50_b32x8_imagenet.py
ShuffleNetV2 1x3x224x224 1.37 727.94 1.19 841.36 1.13 883.47 4.69 213.33 $MMCLS_DIR/configs/shufflenet_v2/shufflenet_v2_1x_b64x16_linearlr_bn_nowd_imagenet.py
MMEditing with 1x3x32x32 input
TensorRT PPLNN
Model Input fp32 fp16 in8 fp16 model config file
latency (ms) FPS latency (ms) FPS latency (ms) FPS latency (ms) FPS
ESRGAN 1x3x32x32 12.64 79.14 12.42 80.50 12.45 80.35 7.67 130.39 $MMEDIT_DIR/configs/restorers/esrgan/esrgan_psnr_x4c64b23g32_g1_1000k_div2k.py
SRCNN 1x3x32x32 0.70 1436.47 0.35 2836.62 0.26 3850.45 0.56 1775.11 $MMEDIT_DIR/configs/restorers/srcnn/srcnn_x4k915_g1_1000k_div2k.py
MMSeg with 1x3x512x1024 input
TensorRT PPLNN
Model Input fp32 fp16 in8 fp16 model config file
latency (ms) FPS latency (ms) FPS latency (ms) FPS latency (ms) FPS
FCN 1x3x512x1024 128.42 7.79 23.97 41.72 18.13 55.15 27.00 37.04 $MMSEG_DIR/configs/fcn/fcn_r50-d8_512x1024_40k_cityscapes.py
PSPNet 1x3x512x1024 119.77 8.35 24.10 41.49 16.33 61.23 27.26 36.69 $MMSEG_DIR/configs/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes.py
DeepLabV3 1x3x512x1024 226.75 4.41 31.80 31.45 19.85 50.38 36.01 27.77 $MMSEG_DIR/configs/deeplabv3/deeplabv3_r50-d8_512x1024_80k_cityscapes.py
DeepLabV3+ 1x3x512x1024 151.25 6.61 47.03 21.26 50.38 26.67 34.80 28.74 $MMSEG_DIR/configs/deeplabv3plus/deeplabv3plus_r50-d8_512x1024_80k_cityscapes.py
MMDet with 1x3x800x1344 input
TensorRT PPLNN
Model Input fp32 fp16 in8 fp16 model config file
latency (ms) FPS latency (ms) FPS latency (ms) FPS latency (ms) FPS
YOLOv3 1x3x800x1344 94.08 10.63 24.90 40.17 24.87 40.21 47.64 20.99 $MMDET_DIR/configs/yolo/yolov3_d53_320_273e_coco.py
SSD-Lite 1x3x800x1344 14.91 67.06 8.92 112.13 8.65 115.63 30.13 33.19 $MMDET_DIR/configs/ssd/ssdlite_mobilenetv2_scratch_600e_coco.py
RetinaNet 1x3x800x1344 97.09 10.30 25.79 38.78 16.88 59.23 38.34 26.08 $MMDET_DIR/configs/retinanet/retinanet_r50_fpn_1x_coco.py
FCOS 1x3x800x1344 84.06 11.90 23.15 43.20 17.68 56.57 - - $MMDET_DIR/configs/fcos/fcos_r50_caffe_fpn_gn-head_1x_coco.py
FSAF 1x3x800x1344 82.96 12.05 21.02 47.58 13.50 74.08 30.41 32.89 $MMDET_DIR/configs/fsaf/fsaf_r50_fpn_1x_coco.py
Faster-RCNN 1x3x800x1344 88.08 11.35 26.52 37.70 19.14 52.23 65.40 15.29 $MMDET_DIR/configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py
Mask-RCNN 1x3x800x1344 320.86 3.12 241.32 4.14 - - 86.80 11.52 $MMDET_DIR/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py
MMOCR
TensorRT PPLNN
Model Input fp32 fp16 in8 fp16 model config file
latency (ms) FPS latency (ms) FPS latency (ms) FPS latency (ms) FPS
DBNet 1x3x640x640 10.70 93.43 5.62 177.78 5.00 199.85 34.84 28.70 $MMOCR_DIR/configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py
CRNN 1x1x32x32 1.93 518.28 1.40 713.88 1.36 736.79 - - $MMOCR_DIR/configs/textrecog/crnn/crnn_academic_dataset.py

Performance benchmark

Users can directly test the performance through how_to_evaluate_a_model.md. And here is the benchmark in our environment.

MMEditing
MMEditing PyTorch ONNX Runtime TensorRT PPLNN
Model Task Metrics(Set5) fp32 fp32 fp32 fp16 int8 fp16 model config file
SRCNN Super Resolution PSNR 28.4316 28.4323 28.4323 28.4286 28.1995 28.4311 $MMEDIT_DIR/configs/restorers/srcnn/srcnn_x4k915_g1_1000k_div2k.py
SSIM 0.8099 0.8097 0.8097 0.8096 0.7934 0.8096
ESRGAN Super Resolution PSNR 28.2700 28.2592 28.2592 - - 28.2624 $MMEDIT_DIR/configs/restorers/esrgan/esrgan_x4c64b23g32_g1_400k_div2k.py
SSIM 0.7778 0.7764 0.7774 - - 0.7765
ESRGAN-PSNR Super Resolution PSNR 30.6428 30.6444 30.6430 - - 27.0426 $MMEDIT_DIR/configs/restorers/esrgan/esrgan_psnr_x4c64b23g32_g1_1000k_div2k.py
SSIM 0.8559 0.8558 0.8558 - - 0.8557
SRGAN Super Resolution PSNR 27.9499 27.9408 27.9408 - - 27.9388 $MMEDIT_DIR/configs/restorers/srresnet_srgan/srgan_x4c64b16_g1_1000k_div2k.pyy
SSIM 0.7846 0.7839 0.7839 - - 0.7839
SRResNet Super Resolution PSNR 30.2252 30.2300 30.2300 - - 30.2294 $MMEDIT_DIR/configs/restorers/srresnet_srgan/msrresnet_x4c64b16_g1_1000k_div2k.py
SSIM 0.8491 0.8488 0.8488 - - 0.8488
Real-ESRNet Super Resolution PSNR 28.0297 27.7016 27.7016 - - 27.7049 $MMEDIT_DIR/configs/restorers/real_esrgan/realesrnet_c64b23g32_12x4_lr2e-4_1000k_df2k_ost.py
SSIM 0.8236 0.8122 0.8122 - - 0.8123
EDSR Super Resolution PSNR 30.2223 30.2214 30.2214 30.2211 30.1383 - $MMEDIT_DIR/configs/restorers/edsr/edsr_x4c64b16_g1_300k_div2k.py
SSIM 0.8500 0.8497 0.8497 0.8497 0.8469 -
MMOCR
MMOCR Pytorch ONNXRuntime TensorRT PPLNN OpenVINO
Model Task Metrics fp32 fp32 fp32 fp16 int8 fp16 fp32 model config file
DBNet* TextDetection recall 0.7310 0.7304 0.7198 0.7179 0.7111 0.7304 0.7309 $MMOCR_DIR/configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py
precision 0.8714 0.8718 0.8677 0.8674 0.8688 0.8718 0.8714
hmean 0.7950 0.7949 0.7868 0.7856 0.7821 0.7949 0.7950
CRNN TextRecognition acc 0.8067 0.8067 0.8067 0.8063 0.8067 - - $MMOCR_DIR/configs/textrecog/crnn/crnn_academic_dataset.py
SAR TextRecognition acc 0.9517 0.9287 - - - - - $MMOCR_DIR/configs/textrecog/sar/sar_r31_parallel_decoder_academic.py
MMSeg
MMSeg Pytorch ONNXRuntime TensorRT PPLNN
Model Metrics fp32 fp32 fp32 fp16 int8 fp16 model config file
FCN mIoU 72.25 - 72.36 72.35 74.19 - $MMSEG_DIR/configs/fcn/fcn_r50-d8_512x1024_40k_cityscapes.py
PSPNet mIoU 78.55 - 78.26 78.24 77.97 - $MMSEG_DIR/configs/pspnet/pspnet_r50-d8_512x1024_80k_cityscapes.py
deeplabv3 mIoU 79.09 - 79.12 79.12 78.96 - $MMSEG_DIR/configs/deeplabv3/deeplabv3_r50-d8_512x1024_40k_cityscapes.py
deeplabv3+ mIoU 79.61 - 79.6 79.6 79.43 - $MMSEG_DIR/configs/deeplabv3plus/deeplabv3plus_r50-d8_512x1024_40k_cityscapes.py
Fast-SCNN mIoU 70.96 - 70.93 70.92 66.0 - $MMSEG_DIR/configs/fastscnn/fast_scnn_lr0.12_8x4_160k_cityscapes.py

Notes

  • As some datasets contains images with various resolutions in codebase like MMDet. The speed benchmark is gained through static configs in MMDeploy, while the performance benchmark is gained through dynamic ones.

  • Some int8 performance benchmarks of TensorRT require nvidia cards with tensor core, or the performance would drop heavily.

  • DBNet uses the interpolate mode nearest in the neck of the model, which TensorRT-7 applies quite different strategy from pytorch. To make the repository compatible with TensorRT-7, we rewrite the neck to use the interpolate mode bilinear which improves final detection performance. To get the matched performance with Pytorch, TensorRT-8+ is recommended, which the interpolate methods are all the same as Pytorch.