Train Adapt Optimize (TAO) Toolkit is a Python-based AI toolkit for customizing purpose-built pre-trained AI models with your own data. TAO Toolkit adapts popular network architectures and backbones for your data, allowing you to train, fine-tune, prune, and export highly optimized and accurate AI models for edge deployment.
The pre-trained models accelerate the AI training process and reduce costs associated with large-scale data collection, labeling, and training from scratch. Transfer learning with pre-trained models can be used for AI applications in smart cities, retail, healthcare, industrial inspection, and more.
Build end-to-end services and solutions for transforming pixels and sensor data to actionable insights using the TAO DeepStream SDK and TensorRT. These models are suitable for object detection, classification, and segmentation.
Image Classification is a popular computer vision technique in which an image is classified into one of the designated classes based on the image features. This model card contains pretrained weights for most of the popular classification models. These weights can be used as a starting point with the classification app in TAO Toolkit to facilitate transfer learning.
FasterViT is a hybrid CNN-ViT-based family of backbone from NVIDIA research that achieves SOTA in ImageNet-1K classification with focus on throughput. This family of backbone leverages Hierarchical Attention (HAT) which decomposes global self-attention with quadratic complexity into a multi-level attention with reduced computational cost. FasterViT architecture was designed with TensorRT in mind, so the model throughput is highly optimized in TensorRT SDK. Use FasterViT when you want to achieve SOTA accuracy on your target dataset while achieving higher throughput compared to other Vision Transformers such as Swin and ConvNext.
The NVImageNet dataset has category names aligned with the original ImageNet-1k category names. The original ImageNet-1K is limited to non-commercial use only, but many recent pretraining techniques show the benefits of pretraining on the ImageNet dataset first and then fine-tuning on downstream tasks. Thus, a model that is pretrained on the original ImageNet dataset might not be allowed to train models for NVIDIA products. The NVImageNet dataset is permitted for commercial uses. The NVImageNet dataset is collected from 84 websites that allow images to be used commercially. It also includes data from Bing image search, which only returns results that are free to share and use commercially.
This model was trained using the classification_pyt
entrypoint in TAO. The training algorithm optimizes the network to minimize cross-entropy loss.
The FasterViT models were trained on the NVImageNet dataset.
The FasterViT models have been evaluated on the ImageNet1K validation dataset.
The key performance indicator is accuracy, following the standard evaluation protocol for image classification. The KPI for the evaluation data are reported below.
Model | Top-1 Accuracy |
---|---|
faster_vit_1_224 | 0.677 |
faster_vit_2_224 | 0.684 |
faster_vit_4_224 | 0.692 |
The inference is run on the provided unpruned model at FP16 precision. The inference performance is run using trtexec
on Jetson AGX Xavier, Xavier NX, Orin, Orin NX, NVIDIA T4, and Ampere GPUs. The Jetson devices are running at Max-N configuration for maximum GPU frequency. The performance shown here is for inference only. End-to-end performance with streaming video data might vary slightly depending on other bottlenecks in the hardware and software.
FasterViT-1-224
Platform | BS | FPS |
---|---|---|
Jetson Orin Nano | 4 | 199 |
Orin NX 16GB | 4 | 292 |
AGX Orin 64GB | 8 | 773 |
A2 | 32 | 761 |
T4 | 32 | 1214 |
A30 | 16 | 3490 |
L4 | 32 | 2871 |
L40 | 32 | 8342 |
A100 | 32 | 5653 |
H100 | 32 | 8825 |
FasterViT-2-224
Platform | BS | FPS |
---|---|---|
Jetson Orin Nano | 4 | 231 |
Orin NX 16GB | 4 | 292 |
AGX Orin 64GB | 8 | 628 |
A2 | 32 | 544 |
T4 | 32 | 889 |
A30 | 16 | 2719 |
L4 | 32 | 1753 |
L40 | 32 | 5607 |
A100 | 32 | 4698 |
H100 | 32 | 7363 |
FasterViT-4-224
Platform | BS | FPS |
---|---|---|
Jetson Orin Nano | 4 | 43 |
Orin NX 16GB | 4 | 59 |
AGX Orin 64GB | 8 | 149 |
A2 | 32 | 117 |
T4 | 32 | 198 |
A30 | 16 | 666 |
L4 | 32 | 421 |
L40 | 32 | 1106 |
A100 | 32 | 1375 |
H100 | 32 | 2510 |
These models must be used with NVIDIA hardware and software. For hardware, the models can run on any NVIDIA GPU including NVIDIA Jetson devices. These models can only be used with TAO Toolkit, the DeepStream SDK, or TensorRT.
The primary use case for these models is classifying objects in a color (RGB) image. They can be used to classify objects from photos and videos by using appropriate video or image decoding and pre-processing.
These models are intended for training and fine-tune using TAO Toolkit and user datasets for image classification. High-fidelity models can be trained to new use cases. A Jupyter Notebook is available as a part of the TAO container and can be used to re-train the models.
The models are also intended for easy edge deployment using DeepStream SDK or TensorRT. DeepStream provides the facilities to create efficient video analytic pipelines to capture, decode, and pre-process the data before running inference.
Category labels (1000 classes) for the input image.
To use these models as pretrained weights for transfer learning, use the following code snippet as a template for the model
and train
component of the experiment spec file when training a FasterViT Classification model. For more information on the experiment spec file, refer to the TAO Toolkit User Guide.
model:
init_cfg:
checkpoint: /path/to/the/faster_vit_0_224.pth
backbone:
type: faster_vit_0_224
head:
type: LinearClsHead
Documentation to deploy with DeepStream is provided in the "Deploying to DeepStream" chapter of the TAO User Guide.
FasterViT was trained on the NVImageNet dataset with 1000 object categories. Hence, the model might not perform well on different data distributions. We recommend conducting further fine-tuning on the target domain to get higher accuracy rates.
The licenses to use this model is covered by the Model EULA. By downloading the unpruned or pruned version of the model, you accept the terms and conditions of these licenses.
NVIDIA platforms and application frameworks enable developers to build a wide array of AI applications. Consider potential algorithmic bias when choosing or creating the models being deployed. Work with the model’s developers to ensure that it meets the requirements for the relevant industry and use case, that the necessary instructions and documentation are provided to understand error rates, confidence intervals, and results, and that the model is being used under the conditions and in the manner intended.