TAO Toolkit Integration with DeepStream
NVIDIA TAO toolkit is a simple, easy-to-use training toolkit that requires minimal coding to create vision AI models using the user’s own data. Using TAO toolkit, users can transfer learn from NVIDIA pre-trained models to create their own model. Users can add new classes to an existing pre-trained model, or they can re-train the model to adapt to their use case. Users can use model pruning capability to reduce the overall size of the model.
Pre-trained models
There are 2 types of pre-trained models that users can start with - purpose-built pre-trained models and meta-architecture vision models. Purpose-built pre-trained models are highly accurate models that are trained on millions of objects for a specific task. The pre-trained weights for meta-architecture vision models merely act as a starting point to build more complex models. These pre-trained weights are trained on Open image dataset and they provide a much better starting point for training versus starting from scratch or starting from random weights. With the latter choice, users can choose from 100+ permutations of model architecture and backbone. See the illustration below.
The purpose-built models are built for high accuracy and performance. These models can be deployed out of the box for applications in smart city or smart places or can also be used to re-train with user’s own data. All models are trained on millions of objects and can achieve more than very high accuracy on our test data. More information about each of these models is available in Computer Vision Model Zoo chapter of TAO toolkit documentation – Computer Vision Model Zoo or in the individual model cards. Typical use cases and some model KPIs are provided in the table below. PeopleNet can be used for detecting and counting people in smart buildings, retail, hospitals, etc. For smart traffic applications, TrafficCamNet and DashCamNet can be used to detect and track vehicles on the road.
TAO toolkit pretrained models - use cases Model Name
Network Architecture
Number of classes
Accuracy
Use case
DetectNet_v2-ResNet18
4
83.50%
Detect and track cars
DetectNet_v2-ResNet18/34
3
84%
People counting, heatmap generation, social distancing.
DetectNet_v2-ResNet18
4
80%
Identify objects from a moving object
DetectNet_v2-ResNet18
1
96%
Detect face in a dark environment with IR camera
ResNet18
20
91%
Classifying car models
ResNet18
6
96%
Classifying type of cars as coupe, sedan, truck, etc
DetectNet_v2-ResNet18
1
98%
Detect License plates on Vehicles
Tuned ResNet18
36(US)/68(CH)
97%(US)/99%(CH)
Recognize characters in License plates. Available in American and Chinese License plates
MaskRCNN-ResNet50
1
85%
Detect and segment people in crowded environment
UNET
3
94.01
Detect people and provide a semantic segmentation mask in an image
Most models trained with TAO toolkit are natively integrated for inference with DeepStream. If the model is integrated, it is supported by the reference deepstream-app. If the model is not natively integrated in the SDK, you can find a reference application on the GitHub repo. See the table below for information on the models supported. For models integrated into deepstream-app, we have provided sample config files for each of the networks. The sample config files are available in the https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/blob/master/deepstream_app_tao_configs/ folder. The table below also lists config files for each model.
Note
Refer README in package /opt/nvidia/deepstream/deepstream/samples/configs/tao_pretrained_models/README.md to obtain TAO toolkit config files and models mentioned in following table.
The TAO toolkit pre-trained models table shows the deployment information of purpose-built pre-trained models.
TAO toolkit pre-trained models in DeepStream Pre-Trained model
DeepStream reference app
Config files
DLA supported
deepstream-app
deepstream_app_source1_trafficcamnet.txt,config_infer_primary_trafficcamnet.txt,labels_trafficnet.txtYes
deepstream-app
deepstream_app_source1_peoplenet.txt,config_infer_primary_peoplenet.txt,labels_peoplenet.txtYes
deepstream-app
deepstream_app_source1_dashcamnet_vehiclemakenet_vehicletypenet.txt,config_infer_primary_dashcamnet.txt,labels_dashcamnet.txtYes
deepstream-app
deepstream_app_source1_faceirnet.txt,config_infer_primary_faceirnet.txt,labels_faceirnet.txtYes
deepstream-app
deepstream_app_source1_dashcamnet_vehiclemakenet_vehicletypenet.txt,config_infer_secondary_vehiclemakenet.txt,labels_vehiclemakenet.txtYes
deepstream-app
deepstream_app_source1_dashcamnet_vehiclemakenet_vehicletypenet.txt,config_infer_secondary_vehicletypenet.txt,labels_vehicletypenet.txtYes
Yes
No
No
No
No
No
No
No
No
No
No
No
No
The TAO toolkit model arch table shows the deployment information of the open model architecture models from TAO toolkit.
TAO toolkit 3.0 model architecture in DeepStream TAO 3.0 Model Architecture
Location
Reference app
Config file
DS Version
YoloV3
DS5.0 onwards
YoloV4
DS5.0 onwards
YoloV4-tiny
DS6.0 onwards
RetinaNet
DS5.0 onwards
SSD
DS5.0 onwards
DSSD
DS5.0 onwards
FasterRCNN
DS5.0 onwards
UNET
DS5.1 onwards
PeopleSegNet
DS5.0 onwards
EfficientDet
DS6.0 onwards
Multi_Task
DS6.0 onwards
YoloV5
DS6.1 onwards
Note
For Yolov3/ReinaNet/SSD/DSSD/FasterRCNN:
cd $DS_TOP/samples/configs/tao_pretrained_models/then editdeepstream_app_source1_detection_models.txt, let config-file under primary-gie section point to the model config files you want, then rundeepstream-app -c deepstream_app_source1_detection_models.txt
For MaskRCNN:
deepstream-app -c deepstream_app_source1_mrcnn.txtLearn more about running the TAO models here: https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps
Except for UNET and License plate recognition (LPR), all models can be deployed in native etlt format. etlt is the TAO export format. For UNET and LPR model, you will need to convert the etlt file to TensorRT engine before running with DeepStream. Use tao-converter to converter etlt to TensorRT engine. tao-converter are very hardware specific, see table below for the appropriate version for your hardware.
TAO converter Platform
Compute
Link
x86 + GPU
CUDA 10.2 / cuDNN 8.0 / TensorRT 7.1
x86 + GPU
CUDA 10.2 / cuDNN 8.0 / TensorRT 7.1
x86 + GPU
CUDA 11.0 / cuDNN 8.0 / TensorRT 7.1
x86 + GPU
CUDA 11.0 / cuDNN 8.0 / TensorRT 7.1
x86 + GPU
CUDA 11.1 / cuDNN 8.0 / TensorRT 7.2
x86 + GPU
CUDA 11.3 / cuDNN 8.0 / TensorRT 8.0
x86 + GPU
CUDA 11.x / cuDNN 8.x / TensorRT 8.2
x86 + GPU
CUDA 11.7 /cuDNN 8.4 / TensorRT 8.4
x86 + GPU
CUDA 11.8 /cuDNN 8.5 / TensorRT 8.5.1.7
x86 + GPU
CUDA 12.1 /cuDNN 8.6 / TensorRT 8.6
Jetson
JetPack 4.4
Jetson
JetPack 4.5
Jetson
Jetpack 4.6
Jetson
Jetpack 5.0.1
Jetson
Jetpack 5.0.2
Jetson
Jetpack 6.0.1
Clara AGX
CUDA 11.1 / CuDNN 8.0.5 / TensorRT 7.2.2
For more information about TAO and how to deploy TAO models, refer to Integrating TAO Models into DeepStream chapter of TAO toolkit user guide.
For more information about deployment of architecture specific models with DeepStream, refer to https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps and https://github.com/NVIDIA-AI-IOT/deepstream_lpr_app GitHub repo.