Skip to content

Latest commit

 

History

History
108 lines (86 loc) · 7.5 KB

README.md

File metadata and controls

108 lines (86 loc) · 7.5 KB
Raw

EfficientDet Jetson Nano TensorRT Benchmarks

Environment

  • Jetson Nano
    • OS
      • JetPack 4.6.1 + TensorRT(22.05) plugin build Linux jetson-desktop 4.9.253-tegra #1 SMP PREEMPT Mon Jul 26 12:13:06 PDT 2021 aarch64 aarch64 aarch64 GNU/Linux
  • Host
    • nvcr.io/nvidia/tensorrt:22.04-py3

How to benchmarks

Export ONNX model.

Latancy

# FP32
$ /usr/src/tensorrt/bin/trtexec --onnx=_PATH_TO_ONNX_ --workspace=1024

# FP16
$ /usr/src/tensorrt/bin/trtexec --onnx=_PATH_TO_ONNX_ --fp16 --workspace=1024

mAP

Models

Results

Latency vs mAP

Jetson Nano

Model name Input size Kind End-to-End Host Latency [ms] Precision
mAP		 Precision
[email protected]		 Precision
[email protected]		 Precision
mAP(small)		 Precision
mAP(medium)		 Precision
mAP(large)		 Recall
AR@1		 Recall
AR@10		 Recall
AR@100		 Recall
AR@100(small)		 Recall
AR@100(medium)		 Recall
AR@100(large)
EfficientDet-lite0 320x320 FP32 46.34 23.93 39.26 24.57 5.65 25.74 43.00 23.11 34.67 36.95 9.45 42.25 61.58
FP16 40.55 23.93 39.26 24.57 5.65 25.74 43.00 23.11 34.67 36.95 9.45 42.25 61.58
EfficientDet-lite1 384x384 FP32 83.62 29.18 46.42 30.04 10.28 32.43 48.16 26.24 40.22 42.69 16.79 48.42 66.22
FP16 74.44 29.17 46.42 30.03 10.29 32.43 48.16 26.24 40.21 42.69 16.81 48.41 66.24
EfficientDet-lite2 448x448 FP32 123.16 32.69 49.99 34.45 13.43 36.65 51.03 28.45 44.06 46.85 22.09 53.35 68.42
FP16 107.33 32.69 50.01 34.43 13.43 36.66 50.99 28.44 44.05 46.84 22.10 53.36 68.38
EfficientDet-lite3 512x512 FP32 217.52 36.46 53.60 38.53 15.45 42.08 55.38 30.98 47.68 50.41 23.07 57.94 73.31
FP16 188.24 36.46 53.61 38.54 15.45 42.08 55.37 30.97 47.67 50.40 23.07 57.95 73.27
EfficientDet-lite3x 640x640 FP32 389.58 40.04 58.67 42.76 21.10 44.31 55.75 32.88 52.01 55.16 33.36 60.84 72.39
FP16 321.11 40.04 58.67 42.77 21.10 44.31 55.75 32.88 52.01 55.16 33.36 60.84 72.39
EfficientDet-lite4 640x640 FP32 512.60 40.84 58.75 43.61 21.55 45.97 58.22 33.26 52.33 55.31 31.58 61.96 74.35
FP16 409.13 40.83 58.75 43.58 21.55 45.95 58.22 33.26 52.34 55.31 31.58 61.96 74.35
EfficientDet-d0 512x512 FP32 134.33 34.12 52.47 36.16 13.12 40.06 53.44 29.28 44.72 47.19 19.66 55.63 69.41
FP16 118.46 34.12 52.47 36.17 13.13 40.10 53.43 29.30 44.73 47.18 19.64 55.64 69.42
EfficientDet-d1 640x640 FP32 294.00 39.94 59.06 42.31 21.09 45.87 57.40 32.51 50.56 53.45 30.71 60.77 72.76
FP16 251.24 39.94 59.05 42.31 21.08 45.87 57.37 32.52 50.56 53.46 30.70 60.75 72.80
EfficientDet-d2 768x768 FP32 464.61 43.22 62.70 46.35 23.70 48.60 60.20 34.24 53.63 56.81 34.49 63.42 74.21
FP16 405.80 43.21 62.70 46.35 23.69 48.59 60.20 34.23 53.63 56.80 34.46 63.43 74.23
EfficientDet-d3 896x896 FP32 879.18 46.84 66.05 50.35 30.16 51.69 62.14 36.36 57.91 61.23 42.15 66.56 77.44
FP16 760.44 46.83 66.04 50.33 30.14 51.66 62.11 36.35 57.91 61.21 42.11 66.53 77.42
SSD MobileNet v2 320x320 FP32 26.96 20.18 34.74 20.12 2.67 17.58 41.43 21.43 32.83 34.76 7.31 35.50 65.62
FP16 24.31 20.18 34.74 20.12 2.67 17.58 41.43 21.43 32.83 34.76 7.31 35.50 65.62
SSD MobileNet V2 FPNLite 320x320 FP32 39.54 21.97 36.95 22.73 5.32 23.39 38.60 21.76 33.62 36.18 8.06 41.97 60.75
FP16 37.99 21.97 36.95 22.73 5.34 23.39 38.60 21.76 33.61 36.18 8.06 41.97 60.75
SSD MobileNet V2 FPNLite 640x640 FP32 138.84 27.92 46.11 29.55 11.97 30.85 40.89 25.65 41.06 44.23 21.99 48.82 62.29
FP16 133.76 27.92 46.11 29.56 11.97 30.85 40.89 25.65 41.06 44.23 21.99 48.82 62.28
SSD ResNet50 V1 FPN 640x640 FP32 735.68 34.15 51.87 37.16 16.52 37.55 49.31 30.27 48.07 51.75 28.53 56.95 71.40
FP16 435.36 34.16 51.87 37.18 16.52 37.56 49.31 30.27 48.06 51.74 28.53 56.94 71.35

Latency mean (ms)

Youtube video

Code

./trt_efficientdet \
  _PATH_TO_TRT_MODEL_FILE \
  --width=MODEL_INPUT_WIDTH \
  --height=MODEL_INPUT_HEIGHT \
  --file=_PATH_TO_INPUT_VIDEO_FILE \
  --score=0.4 \
  --label=_PATH_TO_/tensorrt-examples/models/coco_labels.txt \
  --output=_PATH_TO_OUTPUT_VIDEO_FILE

EfficientDet-lite0 FP16

EfficientDet-lite1 FP16

EfficientDet-lite2 FP16

EfficientDet-lite3 FP16

EfficientDet-lite3x FP16

EfficientDet-lite4 FP16