hyuto / yolo-nas-onnx Goto Github PK

View Code? Open in Web Editor NEW

59.0 3.0 6.0 67.24 MB

Inference YOLO-NAS ONNX model

Home Page: https://hyuto.github.io/yolo-nas-onnx/

License: MIT License

Python 31.47% JavaScript 27.55% HTML 2.22% CSS 3.47% CMake 1.26% C++ 34.03%

object-detection onnx onnxruntime opencv python yolo-nas javascript cpp opencv-dnn

yolo-nas-onnx's Introduction

YOLO-NAS ONNX

Image Source: https://www.pinterest.com/pin/784752303797219490/

Run YOLO-NAS models with ONNX without using Pytorch. Inferencing YOLO-NAS ONNX models with ONNXRUNTIME or OpenCV DNN.

Generate ONNX Model

Generate YOLO-NAS ONNX model without preprocessing and postprocessing within the model. You can convert the model using the following code after installing super_gradients library.

Example: Exporting YOLO-NAS S

from super_gradients.training import models
from super_gradients.common.object_names import Models

model = models.get(Models.YOLO_NAS_S, pretrained_weights="coco")

model.eval()
model.prep_model_for_conversion(input_size=[1, 3, 640, 640])
model.export("yolo_nas_s.onnx", postprocessing=None, preprocessing=None)

Custom Model

To run custom trained YOLO-NAS model in this project you need to generate custom model metadata. Custom model metadata generated from custom-nas-model-metadata.py to provide additional information from torch model.

Usage

python custom-nas-model-metadata.py -m <CHECKPOINT-PATH> \ # Custom trained YOLO-NAS checkpoint path
                                        -t <MODEL-TYPE> \ # Custom trained YOLO-NAS model type
                                        -n <NUM-CLASSES> # Number of classes

After running that it'll generate metadata (json formated) for you

References

https://github.com/Deci-AI/super-gradients

yolo-nas-onnx's People

Contributors

Stargazers

Watchers

Forkers

chaudhryfaisal garethwright davidko3 mattcrossmarc miho raulgimenez

yolo-nas-onnx's Issues

custom-dataset onnx

when i use default coco dataset and convert it into ONNX, after inference i will get true detect result
but when i use custom dataset , convert yolo-nas train weight "pth" into ONNX, after inference i can't get true detect result
so have u tried custom dataset ONNX on your code and the results?
thanks for answering!!!

License

Can you add a license to this repo?

Great work, BTW. I learnt a lot from this repo.

Feature: import int8 model

There is an option to use converted via tensorRt to int8 model?
Or there is other option to quantization to int8?

BTW: grate job and thank for sharing this code :)

YoloNAS with dynamic input shape

I'm not able to convert the model to onnx with dynamic input shape.

from super_gradients.training import models
net = models.get("yolo_nas_s", pretrained_weights="coco")
models.convert_to_onnx(model=net, input_shape=(3,640,640), out_path="yolo_nas_s.onnx")

Is there any flag I can put in the above code?

I tried to convert using the ultralytics package, but without success. Only works with yolov8.

yolo task=detect mode=export model=yolo_nas_s.pt format=onnx dynamic=True

The console stream is logged into /root/sg_logs/console.log
[2023-09-08 04:54:41] INFO - crash_tips_setup.py - Crash tips is enabled. You can set your environment variable to CRASH_HANDLER=FALSE to disable it
Traceback (most recent call last):
  File "/usr/local/bin/yolo", line 8, in <module>
    sys.exit(entrypoint())
  File "/usr/local/lib/python3.8/dist-packages/ultralytics/cfg/__init__.py", line 420, in entrypoint
    model = YOLO(model, task=task)
  File "/usr/local/lib/python3.8/dist-packages/ultralytics/engine/model.py", line 92, in __init__
    self._load(model, task)
  File "/usr/local/lib/python3.8/dist-packages/ultralytics/engine/model.py", line 138, in _load
    self.model, self.ckpt = attempt_load_one_weight(weights)
  File "/usr/local/lib/python3.8/dist-packages/ultralytics/nn/tasks.py", line 589, in attempt_load_one_weight
    args = {**DEFAULT_CFG_DICT, **(ckpt.get('train_args', {}))}  # combine model and default args, preferring model args
  File "/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py", line 1614, in __getattr__
    raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'YoloNAS_S' object has no attribute 'get'

Regards,
Kevin

v8 vs nas

When performing inference on the web using onnx runtime, do you recommend yolov8 or yolo nas? In your experience, which is faster? Which is more accurate? Also, are you using int8 quantization for yolo nas?

C++ Version crash

Hi!, Thanks for your help, I have a problem using custom model, on python version all works, but on c++ version i get

handleNode DNN/ONNX: ERROR during processing node with 2 inputs and 1 outputs: [Squeeze]:(/model/heads/Squeeze_output_0)

I already get metadata, but i cant fix this

Batch mode

Hello!

Is it possible to easily adapt the code to work in batch mode (one input with multiple images)? Perhaps we also need to modify the process of generating the ONNX model?

Thank you!