yolotriton

yolotriton

Go (Golang) gRPC client for YOLO-NAS, YOLO inference using the Triton Inference Server.

Installation

Use go get to install this package:

go get github.com/dev6699/yolotriton

Get YOLO-NAS, YOLO TensorRT model

Export of quantized YOLO model

Install ultralytics

pip install ultralytics

NOTE: Replace yolo12n.pt with your target model

# Export ONNX format then use trtexec to convert
yolo export model=yolo12n.pt format=onnx
trtexec --onnx=yolo12n.onnx --saveEngine=model_repository/yolov12/1/model.plan

NOTE: Inputs/Outputs still remained as FP32 for compatibility reasons.

# export FP32 TensorRT format directly
yolo export model=yolo12n.pt format=engine

# export quantized FP16 TensorRT
yolo export model=yolo12n.pt format=engine half

# export quantized INT8 TensorRT
yolo export model=yolo12n.pt format=engine int8

References:

1. https://docs.nvidia.com/deeplearning/tensorrt/quick-start-guide/index.html
2. https://docs.ultralytics.com/modes/export/#export-formats
3. https://github.com/NVIDIA/TensorRT/tree/master/samples/trtexec

Troubleshooting:

1. Use trtexec --loadEngine=yolo12n.engine to check the engine.
2. Failed to load the exported engine, check Related issue

Convert to FP16 with onnxconverter_common

NOTE: set keep_io_types=True to keep inputs/outputs as FP32, else it will be changed to FP16

import onnx
from onnxconverter_common import float16

# Load original model
model = onnx.load("model.onnx")

model_fp16 = float16.convert_float_to_float16(
    model,
    # keep_io_types=True,
    node_block_list=[]
)

# Save
onnx.save(model_fp16, "model_fp16.onnx")

Export of quantized YOLO-NAS INT8 model

1. Export quantized onnx model

from super_gradients.conversion.conversion_enums import ExportQuantizationMode
from super_gradients.conversion import DetectionOutputFormatMode
from super_gradients.common.object_names import Models
from super_gradients.training import models

# From custom model
# model = models.get(Models.YOLO_NAS_S, num_classes=1, checkpoint_path='ckpt_best.pth')
model = models.get(Models.YOLO_NAS_S, pretrained_weights="coco")
export_result = model.export(
    "yolo_nas_s_int8.onnx",
    output_predictions_format=DetectionOutputFormatMode.BATCH_FORMAT,
    quantization_mode=ExportQuantizationMode.INT8 # or ExportQuantizationMode.FP16
)

print(export_result)

1. Convert to TensorRT with INT8 builder

trtexec --onnx=yolo_nas_s_int8.onnx --saveEngine=yolo_nas_s_int8.plan --int8

References:

1. https://github.com/Deci-AI/super-gradients/blob/b5eb12ccd021ca77e947bf2dde7e84a75489e7ed/documentation/source/models_export.md

Start triton inference server

docker compose up tritonserver

References:

1. https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/model_repository.html

Sample usage

Check cmd/main.go for more details.

• For help

go run cmd/main.go --help

  -b    Run benchmark.
  -i string
        Inference Image. (default "images/1.jpg")
  -m string
        Name of model being served (Required) (default "yolonas")
  -n int
        Number of benchmark run. (default 1)
  -o float
        Intersection over Union (IoU) (default 0.7)
  -p float
        Minimum probability (default 0.5)
  -t string
        Type of model. Available options: [yolonas, yolonasint8, yolofp16, yolofp32] (default "yolonas")
  -u string
        Inference Server URL. (default "tritonserver:8001")
  -x string
        Version of model. Default: Latest Version

• Sample usage with yolonasint8 model

go run cmd/main.go -m yolonasint8 -t yolonasint8 -i images/1.jpg         

1. processing time: 123.027909ms
prediction:  0
class:  dog
confidence: 0.96
bboxes: [ 669 130 1061 563 ]
---------------------
prediction:  1
class:  person
confidence: 0.96
bboxes: [ 440 30 760 541 ]
---------------------
prediction:  2
class:  dog
confidence: 0.93
bboxes: [ 168 83 495 592 ]
---------------------

• Sample usage to get benchmark results

go run cmd/main.go -m yolonasint8 -t yolonasint8 -i images/1.jpg  -b -n 10

1. processing time: 64.253978ms
2. processing time: 51.812457ms
3. processing time: 80.037468ms
4. processing time: 96.73738ms
5. processing time: 87.22928ms
6. processing time: 95.28627ms
7. processing time: 61.609115ms
8. processing time: 87.625844ms
9. processing time: 70.356198ms
10. processing time: 74.130759ms
Avg processing time: 76.93539ms

Results

Input	Ouput