About Resnet10 object detector: which architecture, how to convert the model to onnx

[Hodu]

Hello, @ChrisDing , @DaneLLL , @AastaLLL

Resnet10 object detector: which architecture / how to interpret outputs

Hi, I am trying to run the Resnet10 object detector included with Deepstream 4.0 SDK with the Python TensorRT API to be able to work on a custom tracker prototype. I can’t find any information what type of object detector this model is (SSD, YOLO, Faster RCNN ?), and therefore I haven’t been able to interpret its output vector. Can you please share more detailed information on the included model and how to use it outside of the nvinfer plugin? It’d be also interesting what dataset this net has …

Referring to the above article, I know that the onnx model file used in sample 04 of MMAPI is resnet10 object detector.

https://docs.nvidia.com/jetson/l4t-multimedia/l4t_mm_vid_decode_trt.html
File name: resnet10_dynamic_batch.onnx

Question 1:
In what framework was this model trained before conversion to onnx? (It seems to have been learned in TLT as expected.)

Question 2:

1. If the learning framework is Pytorch, you need to use the torch.onnx.export function to convert it to onnx.

Before
model = network class
I know I need to create a model object from it.
I would like to know the network definition for creating this model object.

2. In case of non-Patorch, can you tell me how to learn the model as a custom dataset and the code to convert it to onnx?

Thank you.

[AastaLLL]

Hi,

TLT model can only be stored in the .etlt format or converted into TensorRT engine with tlt-export.
We don’t have a converter to parse it into ONNX format.

However, the model can be deserialized into TensorRT directly.
So you can still use it within MMAPI by adding a deserializer.

For other frameworks, since ONNX is popular intermediate format.
You can almost find a corresponding parser or tutorial for the frameworks you want to use.

For example:
TensorFlow to ONNX: GitHub - onnx/tensorflow-onnx: Convert TensorFlow models to ONNX
PyTorch to ONNX: Exporting a Model from PyTorch to ONNX and Running it using ONNX Runtime
MXNet to ONNX: Exporting to ONNX format — Apache MXNet documentation

Thanks.

[Hodu]

Hello,

Resnet10_dynamic_batch.onnx provided as a sample
What learning framework (tensorflow, pytorch, etc.) did the original model learn before converting to onnx and converted to onnx?

Thank you.

[DaneLLL]

Hi,
The sample is converted from resnet10.caffemodel. We don’t have public tool for doing it. You may search online to see if there is tool from community. Since the file is for reference only, we suggest convert your model to replace resnet10_dynamic_batch.onnx.

[Hodu]

Hello,

Which data did you use to train the model? If available, can you provide me the code you used to produce the original Caffe model?

Thank you.

[AastaLLL]

Hi,

The model is trained with a private database.
But you should be able to find some similar dataset online.

The model is also trained with a internal tool which similar to the DIGITs.
However, it’s recommended to move to other frameworks since caffe support is much limited now.

For a similar architecture, you can check our Transfer Learning Toolkit.
The output model has its own format (.etlt) and is supported by the deepstream or TensorRT.

NVIDIA Transfer Learning Toolkit

Thanks.