Machine Learning

Stage datasets, map credentials and route outputs for TAO training jobs

Validate an AnomalyGenNext dataset and prepare its texture fine-tuning recipe

Fine-tune CLIP for retrieval and zero-shot tasks and export to ONNX or TensorRT

Fine-tune Cosmos-Embed1 for text-to-video search and semantic deduplication

Fine-tune Cosmos3 models on Cosmos Framework or Cosmos-RL with validated inputs

Fine-tune any Hugging Face vision, VLM or LLM model on local NVIDIA GPUs

Run and fine-tune Kumo-Anomaly diffusion detection on multivariate time series

Run and fine-tune Kumo-Forecast with DARR retrieval on multivariate time series

Fine-tune InternVideo2-CLIP for video-text retrieval and export it to TensorRT

Fine-tune AnomalyGen and generate synthetic anomaly images with quality scoring

Turn a parquet list of image paths into CLIP or SigLIP embedding files

Ground captions to pixel bounding boxes with a two-step VLM labeling pipeline

Generate synthetic defects with AnomalyGenNext and publish COCO labels

Generate referring-expression annotations from images and KITTI boxes with a VLM

Turn raw videos into chain-of-thought training data, captions and QA pairs

Enforce a pre-launch check and run contract for every TAO training or export job

List the TAO models, workflows and AutoML options available in the skill bank

Embed AOI gap targets and mine nearest source images for augmentation

Mine source samples nearest to your targets from embedding parquet files

Mine unique source images for object detection from embedding parquet files

Wire a Hugging Face vision model into TAO config, trainer and TensorRT pipeline

Prepare false negatives, source masks and embeddings as AnomalyGenNext inputs

Split weak Visual ChangeNet samples into per-augmentation subsets by eligibility

Run TAO AutoML hyperparameter search with Bayesian, ASHA, BOHB or LLM methods

Run the three-phase AOI pipeline: AutoML baseline, DEFT loop and refinement

Run the full DEFT loop to improve PCB inspection models with synthetic defects

Run the disk-backed DEFT AOI improvement loop for Cosmos Reason 3 models

Run a mining-based DEFT loop for traffic-camera Cosmos Reason yes or no questions

Run the DEFT augmentation loop to retrain Grounding DINO object detectors

Run a binary inspection DEFT loop with RT-DETR and select models by AP50

Iteratively improve CLIP or SigLIP retrieval on attribute-labelled image data

Select unlabeled images for DINOv3 retraining in an iterative active-learning loop

Start, query and stop a TAO inference microservice in a container

Run a Python training or eval script in an existing virtualenv, no containers

Train, evaluate and export TAO action recognition models on video clips

Train BEVFusion models that fuse LiDAR and camera data for 3D object detection

Train CenterPose models for 6-DoF object pose and keypoint estimation

Train Co-DETR object detection models with hybrid assignment in TAO

Train, quantize and export Deformable DETR models for 2D object detection

Train monocular depth estimation models based on Depth Anything v2

Train DINO object detectors with distillation and quantization in TAO

Domain-adapt DINOv3 vision backbones on unlabeled images with self-supervision

Train and run FastFoundationStereo for real-time stereo depth from image pairs

Train FoundationStereo models that predict disparity from stereo image pairs

Train Grounding DINO for open-set object detection driven by text prompts

Train, distill and quantize PyTorch image classifiers with many backbones

Pretrain and fine-tune Masked Auto-Encoder backbones for vision tasks

Train Mask Auto-Label models that build segmentation masks from point or box hints

Train Mask Grounding DINO for text-prompted instance segmentation