Frequently Asked Questions#

This document includes questions and answers to issues that you might encounter using TAO.

Installation and Usage#

How do I install TAO?

TAO ships as a skill bank plugin for your coding agent (Claude Code, Codex). Install the plugin once, reload it in your session, and export the credentials your chosen compute backend needs. Refer to Getting Started with NVIDIA TAO Toolkit for the full step-by-step.

How do I run training without writing YAML by hand?

Drive the agent in plain English. Describe the model, the dataset URI, the metric you care about, and the backend to use — for example, “Fine-tune DINO on s3://my-bucket/coco-subset/ with 4 GPUs on the local Docker daemon.” The agent reads the relevant SKILL.md files, fills in the spec, shows you the resolved sdk.create_job(...) call, and submits on your confirmation. You never edit spec YAML directly unless you choose to.

Model support#

What are purpose-built models? Can I deploy them in production?

Purpose-built models are highly accurate models trained for applications in smart city, retail, healthcare, and others. NVIDIA trains these production-quality models on very large proprietary datasets for best accuracy and performance.

Are all the models free to use and distribute?

Yes, all models are free to use and distribute. For exact terms for purpose-built models, read our models EULA.

Do I need to re-train the purpose-built models or can I deploy them as is from NGC?

Purpose-built models can be deployed as is using the “pruned” version from the model card but can also be re-trained to better adapt to your environment. For training use the “unpruned” version from the model card.

Instead of NVIDIA provided pretrained models, can I use TAO with my own or any open source pretrained models?

TAO does not support third-party pretrained models. Only NVIDIA pretrained models from NGC are currently supported which can be retrained with your custom data.

Is YOLOv3 supported in TAO?

Yes, YOLOv3 is supported in TAO.

Pruning#

How do I determine the pruning threshold for my model?

The default threshold is 0.1. Every threshold will result in different portions of weights to be pruned, which is reported at the end of the pruning process. A common practice is to prune with increasing threshold values, starting from 0.1 or 0.05. A larger threshold will lead to more weights/channels to be pruned, thus it is harder to restore accuracy or mAP.

Is pruning performed automatically or are there hyperparameters that I need to set to prune my model?

There are multiple parameters for pruning.

  • normalizer is to choose method to normalize weights, default is max

  • equalization_criterion is to choose method to merge weights from different branches of element wise or depth wise layers, default is union

  • pruning_granularity is to set granularity when channels are pruned

  • min_num_filters is to set minimal channels that pruning needs to retain

  • excluded_layers can be used to exclude layers from being pruned

  • pruning_threshold is the most important option. It is used to set the threshold of pruning, which is also used together with Normalizer. This threshold is common for all layers.

Model Export and Deployment#

What is the model output format? How can I use this model for deployment?

TAO model skills export .onnx. To deploy, ask the agent to run the model’s gen_trt_engine action, or run trtexec against the ONNX directly.

How do I deploy models trained with TAO to DeepStream?

Refer to the TAO Quick Start Guide <tao_quick_start_guide> and the deepstream_tlt_apps repository.

Will this model only work with DeepStream? Can I deploy the model without DeepStream?

Deployment to DeepStream is the recommended path for TAO models. The exported ONNX (or a TensorRT engine built from it) can also be deployed outside of DeepStream using TensorRT directly, in which case you are responsible for image pre-processing and output tensor post-processing.

Training#

Is there a dependency of batch size on the accuracy of the model? How should I choose the appropriate batch size for my training?

As a common practice, a small batch size or single GPU is preferred for a small dataset; while a large batch size or multiple GPUs is preferred for a large dataset.

I am seeing lower accuracy with multi-GPU vs. single GPU. Can multi-GPU training affect the accuracy of the model? How do I improve the accuracy in multi-GPU training?

To improve the accuracy in a multi-GPU environment, learning rate parameters need to be higher, for example max_learning_rate. Multi-GPU is preferred only when the training dataset is large.

Distribute the dataset class: How do I balance the weight between classes if the dataset has significantly higher samples for one class versus another?

To account for imbalance, increase the class_weight for classes with fewer samples. You can also try disabling enable_autoweighting; in this case initial_weight is used to control cov/regression weighting. It is important to keep the number of samples of different classes balanced, which helps improve mAP.

How do I save checkpoints in TAO?

TAO saves checkpoints automatically during training. The interval is controlled by the train.checkpoint_interval spec key (default: every 10 epochs for most networks). Or ask the agent: “Save a checkpoint every N epochs.”

In DetectNet_v2, are there any parameters that can help improve AP (average precision) on training small objects?

Following parameters can help you improve AP on smaller objects:

  • Increase num_layers of resnet

  • class_weight for small objects

  • Increase the coverage_radius_x and coverage_radius_y parameters of the bbox_rasterizer_config section for the small objects class

  • Decrease minimum_detection_ground_truth_overlap

  • Lower minimum_height to cover more small objects for evaluation.

Why do I get this error when running tasks in the PyTorch container?

pytorch_lightning.utilities.exceptions.MisconfigurationException:
        you restored a checkpoint with current_epoch=10
        but the Trainer(max_epochs=1)

After you have already trained a model for a number epochs, you can not continue the training by setting the number of epochs (max_epochs) to a lower number than the one already trained for.

Can I run TAO without network?

Pick a backend that runs in your own environment (local Docker, SLURM, Kubernetes) and pre-stage the TAO container images on that environment before dispatching jobs. The agent’s Execution SDK still needs to reach NGC to resolve image tags unless you override to a local registry; see the platform skill for your backend for the exact image-override knobs.