Tao Train Dinov3

NVIDIAOfficialCreated by NVIDIA Corporation

Skill
Tool integration
Data and AI
Apache-2.0
Medium risk

From the author

DINOv3 continual self-supervised pre-training. Domain-adapts public DINOv3 ViT backbones on unlabeled images via teacher-student self-distillation (DINO + iBOT + KoLeo, optional Gram anchoring) and converts the EMA teacher into a timm-format backbone for downstream tasks. Trigger phrases include "train DINOv3", "DINOv3 SSL", "domain-adapt a foundation backbone", "continual pretraining", "self-supervised finetune DINOv3".

— the description by NVIDIA Corporation

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The tao-train-dinov3 skill provides a workflow for domain-adapting DINOv3 vision transformer backbones using self-supervised learning on unlabeled image data. It employs teacher-student self-distillation techniques, including DINO, iBOT, and KoLeo objectives, with optional Gram anchoring. The skill supports full training or parameter-efficient fine-tuning via LoRA, and includes utilities to convert, export, and run inference with trained teacher checkpoints.

What it does

Performs continual self-supervised pre-training on unlabeled image datasets.
Supports teacher-student self-distillation using DINO, iBOT, and KoLeo objectives.
Enables parameter-efficient fine-tuning using LoRA adapters.
Provides tools to convert, export, and perform inference with trained teacher checkpoints.
Supports AutoML for automated hyperparameter optimization.
Allows for high-resolution training configurations.

When to use it

When you need to adapt a pre-trained DINOv3 backbone to a specific domain using unlabeled data.
When you want to improve representation quality for downstream tasks like classification, detection, or segmentation.
When you need to perform continual learning on a foundation model.
When you require a standardized workflow to export or convert DINOv3 models for deployment.

Advantages

Leverages advanced self-supervised objectives for robust feature learning.
Supports LoRA for efficient adaptation without full parameter updates.
Includes built-in support for AutoML to simplify hyperparameter tuning.
Provides clear separation between training, evaluation, and deployment artifacts.
Compatible with standard timm-format backbones.

Usage examples

“Train a DINOv3 backbone on my unlabeled dataset.”: The agent initiates a training run using the specified configuration, applying DINO/iBOT objectives to adapt the backbone to the provided images.
“Export my trained DINOv3 teacher checkpoint for deployment.”: The agent processes the selected teacher checkpoint and converts it into the required format for downstream inference.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/tao-train-dinov3/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip tao-train-dinov3.zip -d .agents/skills

Documentation by NVIDIA Corporation

DINOv3

Standalone install? If this session was not initialized by the TAO skill bank plugin, run the tao-setup skill first.

Use this skill to continue self-supervised training of a DINOv3 backbone on unlabeled images, then convert, export, or run inference with a trained teacher checkpoint.

Quick Start (docker run)

Docker-native launch — no TAO SDK and no Python on the host. Use the local Docker/platform skill instead when it gives a stricter environment-specific command (non-root UID mapping, cache redirects, remote daemons).

TAO_PYT_IMAGE_DEFAULT=nvcr.io/nvidia/tao/tao-toolkit:7.3.0-pyt  # versions-key: images.tao_toolkit.pyt
TAO_PYT_IMAGE="${TAO_PYT_IMAGE:-$TAO_PYT_IMAGE_DEFAULT}"
RUN_ROOT="${RUN_ROOT:-$PWD}"
DOCKER_COMMON=(
  --rm --gpus all --shm-size=8g
  --shm-size=8g
  --ulimit memlock=-1
  --ulimit stack=67108864
  -v "$RUN_ROOT/data:/data:ro"
  -v "$RUN_ROOT/specs:/specs:ro"
  -v "$RUN_ROOT/results:/results"
)

Train:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  dinov3 train -e /specs/train.yaml

Inference:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  dinov3 inference -e /specs/inference.yaml

Export:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  dinov3 export -e /specs/export.yaml

Convert:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  dinov3 convert -e /specs/convert.yaml

Every action takes its spec with -e; results_dir is set in the spec or overridden on the command line. Mount any pretrained-weights directory the spec references, and keep every in-container path consistent across actions.

Configuration

Use schemas/<action>.schema.json for supported fields, defaults, ranges, and options. Start from the matching references/spec_template_<action>.yaml.

Training is AutoML-enabled. Read references/skill_info.yaml and resolve an explicit automl_policy or user request before training. Default to automl_policy: on; requests such as "disable AutoML", "no HPO", or "plain training" mean off for that run. When it is on and the train schema and template are packaged, route training through tao-skill-bank:tao-run-automl. Use direct training only when the policy is off or those files are missing, and report the missing-file limitation. Non-train actions remain in this skill.

Treat train_loss as an optimization signal; use downstream evaluation when selecting a checkpoint for deployment.

For method background or tuning ideas, optionally read:

references/dinov3-method.md
references/dinov3-recipes.md

Core workflow

Obtain a compatible DINOv3 checkpoint in timm or TAO format.
Point dataset.train_dataset.images_dir to a directory of unlabeled images.
Configure the backbone, training scale, and output location.
Run training and evaluate useful teacher checkpoints on data representative of the downstream task.
Use the selected teacher checkpoint with convert, export, or inference.

Provide either train.pretrained_model_path for a new run or train.resume_training_checkpoint_path to resume an existing run.

Essential fields

Purpose	Spec field
Training images	`dataset.train_dataset.images_dir`
Initial weights	`train.pretrained_model_path`
Resume checkpoint	`train.resume_training_checkpoint_path`
Backbone	`model.backbone.teacher_type`, `model.backbone.student_type`
Training scale	`dataset.batch_size`, `train.num_gpus`, `train.num_nodes`
Output directory	`results_dir`
Conversion input	`convert.checkpoint`
Export input	`export.checkpoint`
Inference inputs	`inference.checkpoint`, `dataset.test_dataset.images_dir`

Example train settings:

model:
  backbone:
    teacher_type: vit_b
    student_type: vit_b
    rope_theta: 100.0
dataset:
  train_dataset:
    images_dir: /path/to/unlabeled/images
  batch_size: 16
train:
  pretrained_model_path: /path/to/dinov3/weights
  num_epochs: 10
  num_gpus: 1

Supported image extensions include jpg, jpeg, png, ppm, bmp, pgm, tif, tiff, and webp. Keep evaluation data separate from training data when measuring downstream quality.

Every dataset.*.images_dir value must be an extracted image folder. Managed archive-backed sources use runtime: extracted_folder, which unpacks the source before injecting the directory into the spec. For direct Docker or TAO CLI runs, extract .tar, .tar.gz, .tgz, or .zip inputs before launch.

Checkpoints

File	Use
`model_epoch__step_.pth`	Resume training, or export its EMA teacher
`teacher_epoch__step_.pth`	Convert, export, inference, or downstream evaluation
`student_epoch__step_.pth`	Diagnostics

For export, prefer a selected stripped teacher_epoch_*_step_*.pth. Export also accepts a selected full model_epoch_*_step_*.pth and deterministically selects and logs teacher.backbone. Inference requires a stripped teacher checkpoint. Do not substitute dinov3_model_latest.pth without evaluating that milestone.

Training loss does not directly measure downstream representation quality. When possible, compare candidate teacher checkpoints using metrics relevant to the target task.

For convert, export, and inference, copy the training-time backbone configuration into the action spec, including rope_theta. Stripped checkpoints do not restore the RoPE frequency base, so falling back to the default can change the trained model's behavior.

Key options

model.backbone.teacher_type and student_type support vit_s, vit_s_plus, vit_b, vit_l, vit_h_plus, and vit_7b.
When model.distill.enable: false, the teacher and student backbone types must match.
model.backbone.rope_theta controls the rotary frequency base. It defaults to 100.0 and is tunable; changing it alters positional encoding, so evaluate the resulting checkpoint for the intended task.
model.centering_method supports sinkhorn and softmax.
model.gram.* controls optional Gram anchoring.
model.lora.* enables parameter-efficient continual pre-training. The default targets the attention qkv and proj projections; num_last_blocks: 0 adapts every transformer block.
model.preservation.* adds CLS-token MSE and cosine losses against the frozen anchor teacher. It is recommended with LoRA when global embedding quality matters.
train.log_every_n_steps controls step-level loss visibility and defaults to 1, so short smoke runs emit component and preservation losses.
train.precision defaults to 16-mixed; choose precision based on hardware and measured quality.
train.use_custom_attention: false uses native attention when custom attention is unavailable.
Crop and export dimensions need to be divisible by model.backbone.patch_size.

Distributed training

For iterative GRIT or multi-task weakness mining from an unlabeled pool followed by DINOv3 SSL retraining, use the related tao-run-dinov3-ssl-deft application. Ordinary pretraining and fine-tuning remain in this model skill.

Use train.num_gpus, train.gpu_ids, train.num_nodes, and train.distributed_strategy. Start with auto; use ddp or fsdp when appropriate for the selected backbone, resolution, and available memory.

Common issues

Checkpoint load failure: verify that the checkpoint is DINOv3 and that its backbone matches the configured teacher and student types.
No images found or archive rejected: pass an extracted image folder to dataset.*.images_dir, not an archive file.
Out of memory: reduce batch size or image resolution, or use a sharded distributed strategy.
Custom-attention failure: set train.use_custom_attention: false.
Deterministic-backward failure: xformers memory-efficient attention does not provide a deterministic backward kernel; use train.cudnn.deterministic: false.
Grid-size error: choose crop and export dimensions divisible by the patch size.
Unexpected inference keys: use a stripped teacher_epoch_*_step_*.pth checkpoint.
Full-checkpoint export: a selected model_epoch_*_step_*.pth is supported; export selects and logs its EMA teacher.backbone.
Resume mismatch: the SSL dataloader is not step-resumable, so resume from an epoch/checkpoint boundary when reproducibility matters.

Files

ATTRIBUTION.mdAdded by AutoGineer918 B
BENCHMARK.md7.1 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md8.8 KiB
config/skillspector-baseline.yaml1.0 KiB
evals/evals.json2.2 KiB
references/dinov3-method.md2.3 KiB
references/dinov3-recipes.md2.9 KiB
references/skill_info.yaml3.5 KiB
references/spec_template_convert.yaml2.6 KiB
references/spec_template_export.yaml2.5 KiB
references/spec_template_inference.yaml2.3 KiB
references/spec_template_train.yaml3.9 KiB
references/spec_template_train_highres.yaml3.3 KiB
schemas/convert.schema.json56.0 KiB
schemas/export.schema.json56.9 KiB
schemas/inference.schema.json56.9 KiB
schemas/manifest.json19.8 KiB
schemas/train.schema.json54.1 KiB
skill-card.md4.7 KiB
skill.oms.sig7.8 KiB

schemas/manifest.json

{
  "actions": {
    "convert": {
      "automl_default_parameters": [
        "dataset.batch_size",
        "dataset.workers"
      ],
      "automl_disabled_parameters": [
        "convert",
        "dataset",
        "dataset.test_dataset",
        "dataset.train_dataset",
        "dataset.transform",
        "dataset.transform.global_crops_scale",
        "dataset.transform.local_crops_scale",
        "export",
        "gen_trt_engine",
        "gen_trt_engine.tensorrt",
        "gen_trt_engine.tensorrt.layers_precision",
        "inference",
        "inference.gpu_ids",
        "model",
        "model.backbone",
        "model.distill",
        "model.gram",
        "model.head",
        "model.lora",
        "model.lora.target_modules",
        "model.preservation",
        "train",
        "train.checkpointer",
        "train.cudnn",
        "train.gpu_ids",
        "train.optim",
        "train.schedulers",
        "train.schedulers.last_layer_learning_rate",
        "train.schedulers.learning_rate",
        "train.schedulers.momentum",
        "train.schedulers.teacher_temperature",
        "train.schedulers.weight_decay",
        "wandb",
        "wandb.tags"
      ],
      "core_module": "dinov3",
      "path": "schemas/convert.schema.json",
      "popular": {
        "convert": {
          "checkpoint": "",
          "output_path": "",
          "source": "teacher"
        },
        "dataset": {
          "batch_size": 4,
          "pin_memory": true,
          "transform": {
            "global_crops_scale": [
              0.32,
              1.0
            ],
            "global_crops_size": 256,
            "local_crops_scale": [
              0.05,
              0.32
            ],
            "local_crops_size": 112,
            "n_global_crops": 2,
            "n_local_crops": 8
          },
          "workers": 8
        },
        "gen_trt_engine": {
          "batch_size": -1,
          "gpu_id": 0,
          "tensorrt": {
            "max_batch_size": 1,
            "min_batch_size": 1,
            "opt_batch_size": 1
          }
        },
        "inference": {
          "gpu_ids": [
            0
          ],
          "num_gpus": 1,
          "num_nodes": 1
        },
        "model": {
          "backbone": {
            "drop_path_rate": 0.4,
            "img_size": 256,
            "num_register_tokens": 4,
            "patch_size": 16,
            "rope_theta": 100.0
          },
          "centering_method": "sinkhorn",
          "distill": {
            "disable_masking": false,
            "enable": false
          },
          "gram": {
            "enable": false,
            "start_step": 0,
            "w_gram": 0.0
          },
          "head": {
            "bottleneck_dim": 384,
            "hidden_dim": 2048,
            "num_layers": 3
          },
          "lora": {
            "alpha": 16.0,
            "enable": false,
            "num_last_blocks": 0,
            "rank": 8,
            "target_modules": [
              "qkv",
              "proj"
            ]
          },
          "preservation": {
            "cls_cosine_weight": 0.05,
            "cls_mse_weight": 0.05,
            "enable": false
          }
        },
        "train": {
          "checkpoint_interval": 1,
          "clip_grad_norm": 3.0,
          "distributed_strategy": "auto",
          "gpu_ids": [
            0
          ],
          "layerwise_decay": 1.0,
          "num_epochs": 10,
          "num_gpus": 1,
          "num_nodes": 1,
          "num_prototypes": 131072,
          "optim": {
            "optim": "adamw"
          },
          "precision": "16-mixed",
          "schedulers": {
            "last_layer_learning_rate": {
              "freeze_steps": 1250,
              "max_decay_steps": 2500000,
              "val_base": 7.07e-06,
              "val_final": 1e-06,
              "val_start": 0.0,
              "warm_up_steps": 100000
            },
            "learning_rate": {
              "max_decay_steps": 2500000,
              "val_base": 7.07e-06,
              "val_final": 1e-06,
              "val_start": 0.0,
              "warm_up_steps": 100000
            },
            "momentum": {
              "max_decay_steps": 2500000,
              "val_base": 0.994,
              "val_final": 1.0,
              "val_start": 0.0,
              "warm_up_steps": 0
            },
            "teacher_temperature": {
              "max_decay_steps": 37500,
              "val_base": 0.07,
              "val_final": 0.07,
              "val_start": 0.04,
              "warm_up_steps": 37500
            },
            "weight_decay": {
              "max_decay_steps": 2500000,
              "val_base": 0.04,
              "val_final": 0.2,
              "val_start": 0.0,
              "warm_up_steps": 0
            }
          },
          "use_custom_attention": true,
          "validation_interval": 1
        }
      },
      "schema_action": "convert",
      "spec_template": "references/spec_template_convert.yaml"
    },
    "export": {
      "automl_default_parameters": [
        "dataset.batch_size",
        "dataset.workers"
      ],
      "automl_disabled_parameters": [
        "convert",
        "dataset",
        "dataset.test_dataset",
        "dataset.train_dataset",
        "dataset.transform",
        "dataset.transform.global_crops_scale",
        "dataset.transform.local_crops_scale",
        "export",
        "gen_trt_engine",
        "gen_trt_engine.tensorrt",
        "gen_trt_engine.tensorrt.layers_precision",
        "inference",
        "inference.gpu_ids",
        "model",
        "model.backbone",
        "model.distill",
        "model.gram",
        "model.head",
        "model.lora",
        "model.lora.target_modules",
        "model.preservation",
        "train",
        "train.checkpointer",
        "train.cudnn",
        "train.gpu_ids",
        "train.optim",
        "train.schedulers",
        "train.schedulers.last_layer_learning_rate",
        "train.schedulers.learning_rate",
        "train.schedulers.momentum",
        "train.schedulers.teacher_temperature",
        "train.schedulers.weight_decay",
        "wandb",
        "wandb.tags"
      ],
      "core_module": "dinov3",
      "path": "schemas/export.schema.json",
      "popular": {
        "convert": {
          "checkpoint": "",
          "output_path": "",
          "source": "teacher"
        },
        "dataset": {
          "batch_size": 4,
          "pin_memory": true,
          "transform": {
            "global_crops_scale": [
              0.32,
              1.0
            ],
            "global_crops_size": 256,
            "local_crops_scale": [
              0.05,
              0.32
            ],
            "local_crops_size": 112,
            "n_global_crops": 2,
            "n_local_crops": 8
          },
          "workers": 8
        },
        "gen_trt_engine": {
          "batch_size": -1,
          "gpu_id": 0,
          "tensorrt": {
            "max_batch_size": 1,
            "min_batch_size": 1,
            "opt_batch_size": 1
          }
        },
        "inference": {
          "gpu_ids": [
            0
          ],
          "num_gpus": 1,
          "num_nodes": 1
        },
        "model": {
          "backbone": {
            "drop_path_rate": 0.4,
            "img_size": 256,
            "num_register_tokens": 4,
            "patch_size": 16,
            "rope_theta": 100.0
          },
          "centering_method": "sinkhorn",
          "distill": {
            "disable_masking": false,
            "enable": false
          },
          "gram": {
            "enable": false,
            "start_step": 0,
            "w_gram": 0.0
          },
          "head": {
            "bottleneck_dim": 384,
            "hidden_dim": 2048,
            "num_layers": 3
          },
          "lora": {
            "alpha": 16.0,
            "enable": false,
            "num_last_blocks": 0,
            "rank": 8,
            "target_modules": [
              "qkv",
              "proj"
            ]
          },
          "preservation": {
            "cls_cosine_weight": 0.05,
            "cls_mse_weight": 0.05,
            "enable": false
          }
        },
        "train": {
          "checkpoint_interval": 1,
          "clip_grad_norm": 3.0,
          "distributed_strategy": "auto",
          "gpu_ids": [
            0
          ],
          "layerwise_decay": 1.0,
          "num_epochs": 10,
          "num_gpus": 1,
          "num_nodes": 1,
          "num_prototypes": 131072,
          "optim": {
            "optim": "adamw"
          },
          "precision": "16-mixed",
          "schedulers": {
            "last_layer_learning_rate": {
              "freeze_steps": 1250,
              "max_decay_steps": 2500000,
              "val_base": 7.07e-06,
              "val_final": 1e-06,
              "val_start": 0.0,
              "warm_up_steps": 100000
            },
            "learning_rate": {
              "max_decay_steps": 2500000,
              "val_base": 7.07e-06,
              "val_final": 1e-06,
              "val_start": 0.0,
              "warm_up_steps": 100000
            },
            "momentum": {
              "max_decay_steps": 2500000,
              "val_base": 0.994,
              "val_final": 1.0,
              "val_start": 0.0,
              "warm_up_steps": 0
            },
            "teacher_temperature": {
              "max_decay_steps": 37500,
              "val_base": 0.07,
              "val_final": 0.07,
              "val_start": 0.04,
              "warm_up_steps": 37500
            },
            "weight_decay": {
              "max_decay_steps": 2500000,
              "val_base": 0.04,
              "val_final": 0.2,
              "val_start": 0.0,
              "warm_up_steps": 0
            }
          },
          "use_custom_attention": true,
          "validation_interval": 1
        }
      },
      "schema_action": "export",
      "spec_template": "references/spec_template_export.yaml"
    },
    "inference": {
      "automl_default_parameters": [
        "dataset.batch_size",
        "dataset.workers"
      ],
      "automl_disabled_parameters": [
        "convert",
        "dataset",
        "dataset.test_dataset",
        "dataset.train_dataset",
        "dataset.transform",
        "dataset.transform.global_crops_scale",
        "dataset.transform.local_crops_scale",
        "export",
        "gen_trt_engine",
        "gen_trt_engine.tensorrt",
        "gen_trt_engine.tensorrt.layers_precision",
        "inference",
        "inference.gpu_ids",
        "model",
        "model.backbone",
        "model.distill",
        "model.gram",
        "model.head",
        "model.lora",
        "model.lora.target_modules",
        "model.preservation",
        "train",
        "train.checkpointer",
        "train.cudnn",
        "train.gpu_ids",
        "train.optim",
        "train.schedulers",
        "train.schedulers.last_layer_learning_rate",
        "train.schedulers.learning_rate",
        "train.schedulers.momentum",
        "train.schedulers.teacher_temperature",
        "train.schedulers.weight_decay",
        "wandb",
        "wandb.tags"
      ],
      "core_module": "dinov3",
      "path": "schemas/inference.schema.json",
      "popular": {
        "convert": {
          "checkpoint": "",
          "output_path": "",
          "source": "teacher"
        },
        "dataset": {
          "batch_size": 4,
          "pin_memory": true,
          "transform": {
            "global_crops_scale": [
              0.32,
              1.0
            ],
            "global_crops_size": 256,
            "local_crops_scale": [
              0.05,
              0.32
            ],
            "local_crops_size": 112,
            "n_global_crops": 2,
            "n_local_crops": 8
          },
          "workers": 8
        },
        "gen_trt_engine": {
          "batch_size": -1,
          "gpu_id": 0,
          "tensorrt": {
            "max_batch_size": 1,
            "min_batch_size": 1,
            "opt_batch_size": 1
          }
        },
        "inference": {
          "gpu_ids": [
            0
          ],
          "num_gpus": 1,
          "num_nodes": 1
        },
        "model": {
          "backbone": {
            "drop_path_rate": 0.4,
            "img_size": 256,
            "num_register_tokens": 4,
            "patch_size": 16,
            "rope_theta": 100.0
          },
          "centering_method": "sinkhorn",
          "distill": {
            "disable_masking": false,
            "enable": false
          },
          "gram": {
            "enable": false,
            "start_step": 0,
            "w_gram": 0.0
          },
          "head": {
            "bottleneck_dim": 384,
            "hidden_dim": 2048,
            "num_layers": 3
          },
          "lora": {
            "alpha": 16.0,
            "enable": false,
            "num_last_blocks": 0,
            "rank": 8,
            "target_modules": [
              "qkv",
              "proj"
            ]
          },
          "preservation": {
            "cls_cosine_weight": 0.05,
            "cls_mse_weight": 0.05,
            "enable": false
          }
        },
        "train": {
          "checkpoint_interval": 1,
          "clip_grad_norm": 3.0,
          "distributed_strategy": "auto",
          "gpu_ids": [
            0
          ],
          "layerwise_decay": 1.0,
          "num_epochs": 10,
          "num_gpus": 1,
          "num_nodes": 1,
          "num_prototypes": 131072,
          "optim": {
            "optim": "adamw"
          },
          "precision": "16-mixed",
          "schedulers": {
            "last_layer_learning_rate": {
              "freeze_steps": 1250,
              "max_decay_steps": 2500000,
              "val_base": 7.07e-06,
              "val_final": 1e-06,
              "val_start": 0.0,
              "warm_up_steps": 100000
            },
            "learning_rate": {
              "max_decay_steps": 2500000,
              "val_base": 7.07e-06,
              "val_final": 1e-06,
              "val_start": 0.0,
              "warm_up_steps": 100000
            },
            "momentum": {
              "max_decay_steps": 2500000,
              "val_base": 0.994,
              "val_final": 1.0,
              "val_start": 0.0,
              "warm_up_steps": 0
            },
            "teacher_temperature": {
              "max_decay_steps": 37500,
              "val_base": 0.07,
              "val_final": 0.07,
              "val_start": 0.04,
              "warm_up_steps": 37500
            },
            "weight_decay": {
              "max_decay_steps": 2500000,
              "val_base": 0.04,
              "val_final": 0.2,
              "val_start": 0.0,
              "warm_up_steps": 0
            }
          },
          "use_custom_attention": true,
          "validation_interval": 1
        }
      },
      "schema_action": "inference",
      "spec_template": "references/spec_template_inference.yaml"
    },
    "train": {
      "automl_default_parameters": [
        "dataset.batch_size",
        "dataset.workers"
      ],
      "automl_disabled_parameters": [
        "convert",
        "dataset",
        "dataset.test_dataset",
        "dataset.train_dataset",
        "dataset.transform",
        "dataset.transform.global_crops_scale",
        "dataset.transform.local_crops_scale",
        "export",
        "gen_trt_engine",
        "gen_trt_engine.tensorrt",
        "gen_trt_engine.tensorrt.layers_precision",
        "inference",
        "inference.gpu_ids",
        "model",
        "model.backbone",
        "model.distill",
        "model.gram",
        "model.head",
        "model.lora",
        "model.lora.target_modules",
        "model.preservation",
        "train",
        "train.checkpointer",
        "train.cudnn",
        "train.gpu_ids",
        "train.optim",
        "train.schedulers",
        "train.schedulers.last_layer_learning_rate",
        "train.schedulers.learning_rate",
        "train.schedulers.momentum",
        "train.schedulers.teacher_temperature",
        "train.schedulers.weight_decay",
        "wandb",
        "wandb.tags"
      ],
      "core_module": "dinov3",
      "path": "schemas/train.schema.json",
      "popular": {
        "convert": {
          "checkpoint": "",
          "output_path": "",
          "source": "teacher"
        },
        "dataset": {
          "batch_size": 4,
          "pin_memory": true,
          "transform": {
            "global_crops_scale": [
              0.32,
              1.0
            ],
            "global_crops_size": 256,
            "local_crops_scale": [
              0.05,
              0.32
            ],
            "local_crops_size": 112,
            "n_global_crops": 2,
            "n_local_crops": 8
          },
          "workers": 8
        },
        "gen_trt_engine": {
          "batch_size": -1,
          "gpu_id": 0,
          "tensorrt": {
            "max_batch_size": 1,
            "min_batch_size": 1,
            "opt_batch_size": 1
          }
        },
        "inference": {
          "gpu_ids": [
            0
          ],
          "num_gpus": 1,
          "num_nodes": 1
        },
        "model": {
          "backbone": {
            "drop_path_rate": 0.4,
            "img_size": 256,
            "num_register_tokens": 4,
            "patch_size": 16,
            "rope_theta": 100.0
          },
          "centering_method": "sinkhorn",
          "distill": {
            "disable_masking": false,
            "enable": false
          },
          "gram": {
            "enable": false,
            "start_step": 0,
            "w_gram": 0.0
          },
          "head": {
            "bottleneck_dim": 384,
            "hidden_dim": 2048,
            "num_layers": 3
          },
          "lora": {
            "alpha": 16.0,
            "enable": false,
            "num_last_blocks": 0,
            "rank": 8,
            "target_modules": [
              "qkv",
              "proj"
            ]
          },
          "preservation": {
            "cls_cosine_weight": 0.05,
            "cls_mse_weight": 0.05,
            "enable": false
          }
        },
        "train": {
          "checkpoint_interval": 1,
          "clip_grad_norm": 3.0,
          "distributed_strategy": "auto",
          "gpu_ids": [
            0
          ],
          "layerwise_decay": 1.0,
          "num_epochs": 10,
          "num_gpus": 1,
          "num_nodes": 1,
          "num_prototypes": 131072,
          "optim": {
            "optim": "adamw"
          },
          "precision": "16-mixed",
          "schedulers": {
            "last_layer_learning_rate": {
              "freeze_steps": 1250,
              "max_decay_steps": 2500000,
              "val_base": 7.07e-06,
              "val_final": 1e-06,
              "val_start": 0.0,
              "warm_up_steps": 100000
            },
            "learning_rate": {
              "max_decay_steps": 2500000,
              "val_base": 7.07e-06,
              "val_final": 1e-06,
              "val_start": 0.0,
              "warm_up_steps": 100000
            },
            "momentum": {
              "max_decay_steps": 2500000,
              "val_base": 0.994,
              "val_final": 1.0,
              "val_start": 0.0,
              "warm_up_steps": 0
            },
            "teacher_temperature": {
              "max_decay_steps": 37500,
              "val_base": 0.07,
              "val_final": 0.07,
              "val_start": 0.04,
              "warm_up_steps": 37500
            },
            "weight_decay": {
              "max_decay_steps": 2500000,
              "val_base": 0.04,
              "val_final": 0.2,
              "val_start": 0.0,
              "warm_up_steps": 0
            }
          },
          "use_custom_attention": true,
          "validation_interval": 1
        }
      },
      "schema_action": "train",
      "spec_template": "references/spec_template_train.yaml"
    }
  },
  "automl_enabled": true,
  "failures": {},
  "model": "tao-train-dinov3",
  "network_arch": "dinov3",
  "schema_version": 1
}

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Declares allowed tools, Needs a command-line tool

The author says: Requires docker + nvidia-container-toolkit.

Security

What our scanner noticed· Medium risk

Large base64-encoded blobs

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

NVIDIA Corporation

Publisher

NVIDIA (Official)

Source

NVIDIA/skills · skills/tao-train-dinov3 · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE-APACHE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by NVIDIA Corporation or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download tao-train-dinov3.zip