search

dusty-nv / jetson-inference

Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.
https://developer.nvidia.com/embedded/twodaystoademo
MIT License
7.74k stars 2.97k forks source link

Which framework do you use? #543

Closed elementdl closed 4 years ago

elementdl commented 4 years ago

Hi @dusty-nv ,

I'm new to IA and computer Vision. I trained myself with YOLOv3, with the darknet framework. I just finished your tutoriel with Hello AI world for the Jetson nano, and I was wondering : which framework do I use when I use SSD-MobileNet-V2?

The reason is I'm trying to use the jetson nano for a part of an autonomous car (a prototype so way smaller) by using real-time object detection, and I'm currently thinking about what framework to use on the jetson, what model to use (YoloV3-tiny or SSD-MobileNet-V2).

Thanks for your response,

dusty-nv commented 4 years ago

Hi @elementdl, the SSD-Mobilenet-v2 model from Hello AI World is from TensorFlow and was converted using this tool: https://github.com/AastaNV/TRT_object_detection

elementdl commented 4 years ago

Hi @dusty-nv,

Thank you for your response. I was wondering, the librairy jetson.inference, is it a library directly for the jetson nano or is it a specific a library for the project hello ai world?

Because, in your tutorial it's very easy to chose a model for object detection or recognition, but it's mandatory to chose specific models. Let's say I would like to use the Yolov3-tiny model. Is there some library in python like jetson.infence that allows me to use this model on the jetson nano?

Sincerely,

dusty-nv commented 4 years ago

The jetson.inference library uses TensorRT underneath for accelerated inferencing on Jetson platforms, including Nano/TX1/TX2/Xavier. TensorRT can load models from frameworks trained with caffe, TensorFlow, PyTorch, or models in ONNX format.

That said, there is also typically some pre-/post-processing code required to support the models. Pre-processing typically includes conversion from RGBA to planar NCHW format that the DNN's expect, in addition to mean pixel subtraction. Post-processing is interpretation of the DNN outputs, for example bounding box clustering for detection models.

The pre-/post-processing support isn't included in jetson.inference for YOLO, however there is a TensorRT sample for YOLO included in JetPack 4.3 - you can find it on your Jetson at /usr/src/tensorrt/samples/python/yolov3_onnx