why the inference time on a single image is slow?

[Linxuxin]

Prerequisite

• [X] I have searched the existing and past issues but cannot get the expected help.
• [X] I have read the FAQ documentation but cannot get the expected help.
• [X] The bug has not been fixed in the latest version.

🐞 Describe the bug

I run the ./demo/image_demo.py in the latest mmyolo project by using following config, model and script: config: ./configs/yolov5/voc/yolov5_n-v61_fast_1xb64-50e_voc.py model: https://download.openmmlab.com/mmyolo/v0/yolov5/yolov5_n-v61_fast_1xb64-50e_voc/yolov5_n-v61_fast_1xb64-50e_voc_20221017_234254-f1493430.pth image: ./demo/demo.jpg and found the inference time is slow by adding the testing code in image_demo.py as follows:

    start_time = time.time()
    result = inference_detector(model, file)
    end_time = time.time()
    print("time cost:", float(end_time - start_time) * 1000.0, "ms")

script: python ./demo/image_demo.py ./demo/demo.jpg ./configs/yolov5/voc/yolov5_n-v61_fast_1xb64-50e_voc.py ./model/yolov5_n-v61_fast_1xb64-50e_voc_20221017_234254-f1493430.pth The result is: time cost: 1094.2001342773438 ms

I compare another yolov5 project: https://github.com/ultralytics/yolov5 in the same environment and run the detect.py by using following model and script: model: https://github.com/ultralytics/yolov5/releases/download/v7.0/yolov5n.pt script: python ./detect.py --weights yolov5n.pt --img 640 --conf 0.25 --source data/images

data/images only contains a demo.jpg provided in mmyolo/demo

The result is: time cost: 4.645109176635742 ms

Please help me and point out where I went wrong.

Environment

sys.platform: linux Python: 3.8.16 (default, Mar 2 2023, 03:21:46) [GCC 11.2.0] CUDA available: True numpy_random_seed: 2147483648 GPU 0,1: NVIDIA TITAN V CUDA_HOME: /usr/local/cuda NVCC: Cuda compilation tools, release 11.3, V11.3.109 GCC: gcc (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0 PyTorch: 1.12.1 PyTorch compiling details: PyTorch built with:

• GCC 9.3
• C++ Version: 201402
• Intel(R) oneAPI Math Kernel Library Version 2021.4-Product Build 20210904 for Intel(R) 64 architecture applications
• Intel(R) MKL-DNN v2.6.0 (Git Hash 52b5f107dd9cf10910aaa19cb47f3abf9b349815)
• OpenMP 201511 (a.k.a. OpenMP 4.5)
• LAPACK is enabled (usually provided by MKL)
• NNPACK is enabled
• CPU capability usage: AVX2
• CUDA Runtime 11.3
• NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_37,code=compute_37
• CuDNN 8.3.2 (built against CUDA 11.5)
• Magma 2.5.2
• Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.3, CUDNN_VERSION=8.3.2, CXX_COMPILER=/opt/rh/devtoolset-9/root/usr/bin/c++, CXX_FLAGS= -fabi-version=11 -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -DEDGE_PROFILER_USE_KINETO -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-unused-parameter -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Werror=cast-function-type -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.12.1, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=OFF, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=OFF,

TorchVision: 0.13.1 OpenCV: 4.7.0 MMEngine: 0.7.2 MMCV: 2.0.0 MMDetection: 3.0.0 MMYOLO: 0.5.0+dc85144

Additional information

No response

[hhaAndroid]

@Linxuxin This is reasonable, because a low post-processing threshold will result in a lot of output bbox

[Linxuxin]

@hhaAndroid Thanks a lot for your reply. I check my scripts and reuse the following model and default config to test the run time:

python ./demo/image_demo.py ./demo/demo.jpg ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py ./model/yolov5_n-v61_syncbn_fast_8xb16-300e_coco_20220919_090739-b804c1ad.pth time cost: 1114.0129566192627 ms

Based on the above model, I use the easydeploy tool to change the post-processing threshold and test again as follows:

1. convert to onnx file python ./projects/easydeploy/tools/export.py ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py ./model/yolov5_n-v61_syncbn_fast_8xb16-300e_coco_20220919_090739-b804c1ad.pth --work-dir work_dirs/yolov5_n-v61_syncbn_fast_8xb16-300e_coco --img-size 640 640 --batch 1 --device cuda:0 --simplify --opset 11 --backend 1 --pre-topk 10 --keep-topk 10 --iou-threshold 0.65 --score-threshold 0.25
2. run the model python projects/easydeploy/tools/image-demo.py ./demo/demo.jpg ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py work_dirs/yolov5_n-v61_syncbn_fast_8xb16-300e_coco/end2end.onnx --device cuda:0 time cost: 49.04508590698242 ms

Here, the inference time is obviously improved. However, there is still a no small gap comparing to the project: https://github.com/ultralytics/yolov5: python ./detect.py --weights yolov5n.pt --img 640 --conf 0.25 --source data/images time cost: 4.645109176635742 ms

Can you give me some explain or advice?

[KNzzzzzzzzzz]

Can you explain now？

@hhaAndroid Thanks a lot for your reply. I check my scripts and reuse the following model and default config to test the run time:

python ./demo/image_demo.py ./demo/demo.jpg ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py ./model/yolov5_n-v61_syncbn_fast_8xb16-300e_coco_20220919_090739-b804c1ad.pth time cost: 1114.0129566192627 ms

Based on the above model, I use the easydeploy tool to change the post-processing threshold and test again as follows:

1. convert to onnx file python ./projects/easydeploy/tools/export.py ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py ./model/yolov5_n-v61_syncbn_fast_8xb16-300e_coco_20220919_090739-b804c1ad.pth --work-dir work_dirs/yolov5_n-v61_syncbn_fast_8xb16-300e_coco --img-size 640 640 --batch 1 --device cuda:0 --simplify --opset 11 --backend 1 --pre-topk 10 --keep-topk 10 --iou-threshold 0.65 --score-threshold 0.25
2. run the model python projects/easydeploy/tools/image-demo.py ./demo/demo.jpg ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py work_dirs/yolov5_n-v61_syncbn_fast_8xb16-300e_coco/end2end.onnx --device cuda:0 time cost: 49.04508590698242 ms

Here, the inference time is obviously improved. However, there is still a no small gap comparing to the project: https://github.com/ultralytics/yolov5: python ./detect.py --weights yolov5n.pt --img 640 --conf 0.25 --source data/images time cost: 4.645109176635742 ms

Can you give me some explain or advice?

[imangoa]

@hhaAndroid Thanks a lot for your reply. I check my scripts and reuse the following model and default config to test the run time:

python ./demo/image_demo.py ./demo/demo.jpg ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py ./model/yolov5_n-v61_syncbn_fast_8xb16-300e_coco_20220919_090739-b804c1ad.pth time cost: 1114.0129566192627 ms

Based on the above model, I use the easydeploy tool to change the post-processing threshold and test again as follows:

1. convert to onnx file python ./projects/easydeploy/tools/export.py ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py ./model/yolov5_n-v61_syncbn_fast_8xb16-300e_coco_20220919_090739-b804c1ad.pth --work-dir work_dirs/yolov5_n-v61_syncbn_fast_8xb16-300e_coco --img-size 640 640 --batch 1 --device cuda:0 --simplify --opset 11 --backend 1 --pre-topk 10 --keep-topk 10 --iou-threshold 0.65 --score-threshold 0.25
2. run the model python projects/easydeploy/tools/image-demo.py ./demo/demo.jpg ./configs/yolov5/yolov5_n-v61_syncbn_fast_8xb16-300e_coco.py work_dirs/yolov5_n-v61_syncbn_fast_8xb16-300e_coco/end2end.onnx --device cuda:0 time cost: 49.04508590698242 ms

Here, the inference time is obviously improved. However, there is still a no small gap comparing to the project: https://github.com/ultralytics/yolov5: python ./detect.py --weights yolov5n.pt --img 640 --conf 0.25 --source data/images time cost: 4.645109176635742 ms

Can you give me some explain or advice?

yolov5 project uses multithread mode to read images，but mmyolo not do it.