Benchmarks Targeted for Jetson (Using GPU+2DLA)
The script will run following Benchmarks:
- Names : Input Image Resolution
- Inception V4 : 299x299
- ResNet-50 : 224x224
- OpenPose : 256x456
- VGG-19 : 224x224
- YOLO-V3 : 608x608
- Super Resolution : 481x321
- Unet : 256x256
For benchmark results on all NVIDIA Jetson Products; please have a look at NVIDIA jetson_benchmark webpage
Following scripts are included:
- Installation requirements for running benchmark script (install_requirements.sh)
- CSV files containing parameters (benchmark_csv folder)
- Download Model (utils/download_models.py)
- Running Benchmark Script (benchmarks.py)
Version Dependencies:
- JetPack 4.4+
- TensorRT 7+
Set up instructions
git clone https://github.com/NVIDIA-AI-IOT/jetson_benchmarks.git
cd jetson_benchmarks
mkdir models # Open folder to store models (Optional)
Install Requirements
sudo sh install_requirements.sh
Note: All libraries will be installed for python3
For Jetson Xavier NX
Download Models
python3 utils/download_models.py --all --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --save_dir <absolute-path-to-downloaded-models>
Running Benchmarks
Running All Benchmark Models at Once
sudo python3 benchmark.py --all --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models>
Sample Output
| Model Name | FPS |
|---|---|
| inception_v4 | 311.73 |
| vgg19_N2 | 66.43 |
| super_resolution_bsd500 | 150.46 |
| unet-segmentation | 145.42 |
| pose_estimation | 237.1 |
| yolov3-tiny-416 | 546.69 |
| ResNet50_224x224 | 824.02 |
| ssd-mobilenet-v1 | 887.6 |
Running an Individual Benchmark Model
-
For Inception V4
sudo python3 benchmark.py --model_name inception_v4 --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models> -
For VGG19
sudo python3 benchmark.py --model_name vgg19 --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models> -
For Super Resolution
sudo python3 benchmark.py --model_name super_resolution --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models> -
For UNET Segmentation
sudo python3 benchmark.py --model_name unet --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models> -
For Pose Estimation
sudo python3 benchmark.py --model_name pose_estimation --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models> -
For Tiny-YOLO-V3
sudo python3 benchmark.py --model_name tiny-yolov3 --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models> -
For ResNet-50
sudo python3 benchmark.py --model_name resnet --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models> -
For SSD-MobileNet-V1 Segmentation
sudo python3 benchmark.py --model_name ssd-mobilenet-v1 --csv_file_path <path-to>/benchmark_csv/nx-benchmarks.csv --model_dir <absolute-path-to-downloaded-models>
For Jetson AGX Xavier
Please follow setup and installation requirements.
Download Models
python3 utils/download_models.py --all --csv_file_path <path-to>/benchmark_csv/xavier-benchmarks.csv --save_dir <absolute-path-to-downloaded-models>
Running All Benchmark Models at Once on Jetson AGX Xavier
sudo python3 benchmark.py --all --csv_file_path <path-to>/benchmark_csv/xavier-benchmarks.csv \
--model_dir <absolute-path-to-downloaded-models> \
--jetson_devkit xavier \
--gpu_freq 1377000000 --dla_freq 1395200000 --power_mode 0 --jetson_clocks
For Jetson TX2 and Jetson Nano
Please follow setup and installation requirements.
Download Models
python3 utils/download_models.py --all --csv_file_path <path-to>/benchmark_csv/tx2-nano-benchmarks.csv --save_dir <absolute-path-to-downloaded-models>
Running All Benchmark Models at Once on Jetson TX2
sudo python3 benchmark.py --all --csv_file_path <path-to>/benchmark_csv/tx2-nano-benchmarks.csv \
--model_dir <absolute-path-to-downloaded-models> \
--jetson_devkit tx2 \
--gpu_freq 1122000000 --power_mode 3 --precision fp16
Running All Benchmark Models at Once on Jetson Nano
sudo python3 benchmark.py --all --csv_file_path <path-to>/benchmark_csv/tx2-nano-benchmarks.csv \
--model_dir <absolute-path-to-downloaded-models> \
--jetson_devkit nano \
--gpu_freq 921600000 --power_mode 0 --precision fp16
For Jetson Orin
Please follow setup and installation requirements.
Download Models
python3 utils/download_models.py --all --csv_file_path <path-to>/benchmark_csv/orin-benchmarks.csv --save_dir <absolute-path-to-downloaded-models>
Running All Benchmark Models at Once on Orin
Please check if Orin is in MAX power mode:
nvpmodel -q
It should say:
NV Power Mode: MAXN
0
If it isn't, set it to MAX power mode:
sudo nvpmodel -m 0. Then run:
sudo python3 benchmark.py --all --csv_file_path <path-to>/benchmark_csv/orin-benchmarks.csv \
--model_dir <absolute-path-to-downloaded-models>
For Jetson Orin Nano
Please follow setup and installation requirements.
Download Models
python3 utils/download_models.py --all --csv_file_path <path-to>/benchmark_csv/orin-nano-benchmarks.csv --save_dir <absolute-path-to-downloaded-models>
Running All Benchmark Models at Once on Orin Nano
sudo python3 benchmark.py --all --csv_file_path <path-to>/benchmark_csv/orin-nano-benchmarks.csv \
--model_dir <absolute-path-to-downloaded-models> --jetson_clocks
For Jetson Orin NX
Please follow setup and installation requirements.
Download Models for Orin NX 8GB
python3 utils/download_models.py --all --csv_file_path <path-to>/benchmark_csv/orin-nx-8gb-benchmarks.csv --save_dir <absolute-path-to-downloaded-models>
Running All Benchmark Models at Once on Orin NX 8GB
sudo python3 benchmark.py --all --csv_file_path <path-to>/benchmark_csv/orin-nx-8gb-benchmarks.csv \
--model_dir <absolute-path-to-downloaded-models>
Download Models for Orin NX 16GB
python3 utils/download_models.py --all --csv_file_path <path-to>/benchmark_csv/orin-nx-16gb-benchmarks.csv --save_dir <absolute-path-to-downloaded-models>
Running All Benchmark Models at Once on Orin NX 16GB
sudo python3 benchmark.py --all --csv_file_path <path-to>/benchmark_csv/orin-nx-16gb-benchmarks.csv \
--model_dir <absolute-path-to-downloaded-models>