AutoGineer
Generated banner for Tao Port Huggingface Model
Generated logo for Tao Port Huggingface Model

Tao Port Huggingface Model

NVIDIAOfficialCreated by NVIDIA Corporation

  • Skill
  • Workflow
  • Software development
  • Apache-2.0
  • Medium risk
Download ZIP

From the author

Integrate a HuggingFace Computer Vision model into the NVIDIA TAO Toolkit ecosystem (tao-core config, tao-pytorch trainer, tao-deploy TensorRT pipeline). Use when the user asks to "integrate a HuggingFace model into TAO", "add an HF model to TAO Toolkit", "wire a HuggingFace ViT/DETR/ SegFormer into tao-pytorch", "build a TAO trainer + deploy pipeline for an HF CV model", or pastes a HuggingFace model URL/ID and wants it turned into a TAO model. Covers the full 7-phase loop: prerequisites check, HuggingFace inspection and validation, codebase exploration, tao-core configuration and native trainer implementation, ONNX export plus TensorRT deploy integration, packaging and L0 testing, container-based end-to-end validation, and (conditional) accuracy/latency tuning. Supports classification, object detection, semantic / instance / panoptic segmentation, zero-shot detection, and depth estimation.
— the description by NVIDIA Corporation

AutoGineer summary

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

This skill provides a structured, seven-phase workflow for porting HuggingFace Computer Vision models into the NVIDIA TAO Toolkit. It guides the user through inspecting the model, configuring the TAO core, implementing the trainer, and setting up the deployment pipeline for TensorRT. The process emphasizes a local-only development loop, requiring the user to manage four independent repository clones and execute all validation, testing, and training tasks within locally prepared TAO Toolkit containers.

What it does

  • Validates HuggingFace model compatibility for computer vision tasks.
  • Configures TAO core and implements native trainers for PyTorch.
  • Integrates ONNX export and TensorRT deployment pipelines.
  • Automates packaging and L0 unit testing for model components.
  • Performs end-to-end validation within containerized environments.
  • Provides optional tuning for accuracy, latency, and model size.

When to use it

  • When you need to bring a custom HuggingFace computer vision model into the TAO Toolkit.
  • When you require a full training-to-deployment pipeline for a new model architecture.
  • When you need to verify model performance and exportability to TensorRT.
  • When you are working in a local environment with NVIDIA GPU support.

Advantages

  • Ensures consistency across training, export, and deployment stages.
  • Provides a rigorous, step-by-step debugging and testing loop.
  • Maintains environment isolation using containerized workflows.
  • Supports a wide range of computer vision tasks including detection and segmentation.

Usage examples

“Integrate a HuggingFace model into TAO.”
The agent initiates the Phase 0 prerequisite check to verify your environment and prepare the necessary TAO Toolkit container images.
“Wire a HuggingFace SegFormer into tao-pytorch.”
The agent begins the codebase exploration phase to map the model to the appropriate TAO reference architecture and prepare the configuration.

Install

  1. Download the skill as a ZIP:

    https://mp-staging.autogineer.com/agent-skills/nvidia/tao-port-huggingface-model/download

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

    mkdir -p .agents/skills && unzip tao-port-huggingface-model.zip -d .agents/skills

Documentation by NVIDIA Corporation

Standalone install? If this session was not initialized by the TAO skill bank plugin, run the tao-setup skill first (host preflight, credentials, cross-skill discovery).

<!-- Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->

TAO-HF Integration Skill

Integrate a HuggingFace (HF) Computer Vision model into the NVIDIA TAO Toolkit ecosystem. Work the phases iteratively — not purely linearly — via a build → test → debug → fix → retest loop at every step: when something fails, diagnose and fix before moving on; when it passes, move to the next step.

This SKILL.md is the workflow coordinator. Each phase has a dedicated references/phase-N-*.md with the full step-by-step content, code, docker invocations, and gates. Read the matching reference at the start of each phase — the summaries below are not sufficient.

Local-Only Rule

All work is strictly local. Do NOT push/commit/branch on any remote (GitLab, GitHub, HuggingFace), create merge/pull requests or issues, or upload/publish Docker images to any registry or artifact store. You may only read/clone from remotes — all edits, Docker builds, and test runs stay on the local machine.

Submodule Override Strategy

The user clones the four TAO repos (tao-core, tao-pytorch, tao-deploy, tao-dataservices) independently into one working directory. The tao-core/ submodule nested inside each repo points to the original unmodified commit; modifications only exist in the top-level tao-core/. Always install from the top-level tao-core/, never <repo>/tao-core/ — the nested submodule silently ignores all modifications. Override rules: (1) mount the working directory -v $(pwd):/workspace; (2) pip install /workspace/tao-core FIRST, before tao-pytorch/tao-deploy; (3) PYTHONPATH top-level tao-core first, e.g. -e PYTHONPATH=/workspace/tao-core:/workspace/tao-pytorch. See references/cross-cutting.md for the directory tree.

Execution platform

Every test, smoke run, and end-to-end validation executes inside a locally prepared TAO Toolkit container (tao-pytorch-base:latest, tao-deploy-base:latest, optionally tao-dataservices-base:latest — all from Phase 0). The platform skills own how to run them; this skill specifies what. Default platform: local-docker. Phase 0 delegates the driver / CUDA / NCT preflight to tao-setup-nvidia-gpu-host. See references/cross-cutting.md for the authoritative-skill table, bind-mount rationale, and canonical docker-run flag set.

Phase Map

PhaseGoalReference
0Prerequisites + TAO Toolkit images + local image tagsphase-0-prereqs.md
1Inputs, HF-inspection container, validate model + datasetphase-1-inspection.md, hf-inspection.md
2Closest existing TAO reference modelphase-2-codebase.md, task-type-guide.md
3tao-core config + tao-pytorch trainer / eval / inferencephase-3-implementation.md, tao-patterns.md, repo-structure.md
4ONNX export + tao-deploy TRT engine / inference / evalphase-4-deploy.md
5Packaging (console_scripts) + L0 testsphase-5-packaging.md
6Container testing + end-to-end validationphase-6-container-tests.md, docker-patterns.md
7(conditional) Accuracy / latency / size tuningphase-7-optimization.md

Cross-cutting refs: workflow-consistency.md (CLI flow, config field paths, cross-phase dependencies); cross-cutting.md (platform, isolation, module pitfalls, debugging).

IMPORTANT — Continuous Execution Through Phase 6: do NOT stop after Phases 3–5 to wait for the user to run tests. Phase 6 is mandatory — not complete until tests pass inside the containers and the end-to-end pipeline is validated.

Development Loop

At every step: write code → test immediately (import check, unit test, or dry-run) → if it fails, read traceback → diagnose → fix → retest; if it passes, move on. Do NOT accumulate untested code — testing only at the end compounds bugs.

Debugging Playbook

When something fails, consult the symptom → likely-cause → fix table in references/cross-cutting.md before trying random fixes — it covers ModuleNotFoundError, BACKBONE_REGISTRY KeyError, shape mismatch, NaN loss, ONNX/TRT build failures, TRT-vs-PyTorch accuracy gaps, OOM, DDP hangs, checkpoint load failures, and stale-submodule config issues.

Environment Isolation Strategy

All Python work runs inside Docker containers — no host venvs, no pip installs into host Python (the host needs only Docker, from tao-setup-nvidia-gpu-host). Three contexts: A (Phase 1 HF inspection in tao-hf-inspect, python:3.12-slim fallback), B (Phase 3/4/6 smoke/L0/e2e in the prepared container, source via pip install /workspace/tao-core && python setup.py develop), C (host-bind-mount scratch). See references/cross-cutting.md for the contexts in full and the four numbered rules verbatim (--check-only host packages; Phase 1 --user $(id -u):$(id -g) vs. root; HOME=/workspace/PIP_USER=1 fallback; distro package-manager list; root:root trade-off; tao-hf-inspect cleanup).

Phase 0 — Prerequisites Check

Goal: verify Python 3.10+ and git; delegate the driver / CUDA / Docker / NVIDIA Container Toolkit host check to tao-setup-nvidia-gpu-host; verify NGC docker login for nvcr.io. Then ask the user for the TAO Toolkit image references (tao-pytorch, tao-deploy, optionally tao-dataservices), pull, and prepare local tags tao-pytorch-base:latest, tao-deploy-base:latest, tao-dataservices-base:latest for later phases — preparation removes the pre-installed released TAO packages so the user's /workspace/... clones install/load via pip install /workspace/tao-core && python setup.py develop. Hard stop on any failed check. Required user inputs: the image references + credentials (NGC login, HF_TOKEN). Full commands, prompt wording, and per-image Dockerfile snippets: the Phase 0 reference.

Gate: all prerequisite checks pass; the user supplied the required image references; tao-pytorch-base:latest and tao-deploy-base:latest exist locally; tao-dataservices-base:latest exists if dataservices work is anticipated.

Phase 1 — Information Gathering & Validation

Goal: decide whether to proceed at all. Gather credentials, locate/clone the four TAO repos, create a consistent working branch, launch the tao-hf-inspect container (Context A), validate the HF model is CV with a supported pipeline_tag, extract config + state-dict schema, sanity-check ONNX export, clean up. Full steps: Phase 1 references.

Reject if: pipeline_tag is NLP / audio / LLM (non-CV); AutoConfig raises; or ONNX export fundamentally cannot work (no rewrite path).

Gate: all 4 TAO repos located/cloned with a consistent branch; pipeline_tag confirmed CV; model_type, image_size, hidden_size, num_labels extracted; state-dict keys documented + HF→TAO remapping plan drafted; ONNX export sanity check passed (or failure understood); user confirmed model_short_name + task type. (Full checklist: phase-1-inspection.md.) Present findings and get user confirmation first.

Phase 2 — Codebase Exploration

Goal: find the closest existing TAO reference model for the detected pipeline_tag, read its implementation across tao-core / tao-pytorch / tao-deploy, and decide whether the backbone exists in backbone_v2/ or is new.

The HF pipeline_tag → TAO reference model mapping (classification → classification_pyt, detection → dino/rtdetr, segmentation → segformer, instance → mask2former, panoptic → oneformer, zero-shot → grounding_dino, depth → mono_depth) drives everything downstream (config, architecture, loss, ONNX shape, TRT builder, deploy classes, metrics, dataset format). See the Phase 2 references for the full reference list (12 files per model), the backbone_v2/ and tao-dataservices coverage checks, and per-task architecture.

If a new backbone is needed, decide the strategy (timm wrap > re-implement > HF black-box wrap) before Phase 3 — it changes weight loading, ONNX export, deploy. Never dual-inherit from transformers.PreTrainedModel and BackboneBase (metaclass conflict — compose instead).

Gate: reference TAO model identified + all 12 reference locations read; task-type implications understood (architecture, loss, ONNX outputs, deploy classes, metrics, dataset); backbone coverage decided (reuse / wrap timm / new); dataservices coverage checked. (Full checklist: phase-2-codebase.md.)

Phase 3 — TAO Core Configuration & Native Implementation

Goal: write the tao-core config schema + the tao-pytorch trainer / native inference / evaluation, smoke-testing between steps. (<model_name> = snake_case short-name; <ModelName> = PascalCase.)

Steps 1–7 (each builds on the previous, smoke-test between): tao-core config (1), tao-pytorch trainer (2), multi-GPU/multi-node (3), native inference → result.csv (4), native evaluation → results.json (5), MLOps for training and eval/infer → status.json (6–7). The ExperimentConfig(CommonExperimentConfig) must contain model, dataset, train, evaluate, inference, export, gen_trt_engine, quantize. All ??? fields are MISSING (user supplies via YAML/CLI); the augmentation.mean/std, model.head.in_channels, checkpoint-name, and onnx_file matches are in the checklist below.

Full per-step bodies, code, the canonical experiment_spec.yaml, and smoke-test commands: the Phase 3 references.

Gates: Step 1 — ExperimentConfig imports cleanly in-container; Step 2 — build_model(cfg) runs + PLModel instantiates in-container; Phase 3 — all 7 steps complete, smoke tests pass, no missing __init__.py.

Phase 4 — Export, Deployment & TensorRT Integration

Goal: ONNX export from tao-pytorch, then TRT engine builder + inference + evaluation in tao-deploy reusing the tao-core ExperimentConfig.

Steps 8–11: ONNX exporter (8 — task-specific input/output names, batch_size=-1 ⇒ dynamic batch); TRT engine builder (9 — subclass EngineBuilder or reuse ClassificationEngineBuilder; write specs/{gen_trt_engine,inference,evaluate}.yaml, same ExperimentConfig schema, augmentation.mean/std MUST match training); TRT inference → result.csv (10); TRT eval → results.json (11). See the Phase 4 reference for full code and the Phase 3+4 gate (3 in-container checks: imports, model build + forward, ONNX round-trip).

Module pitfalls: tao-pytorch and tao-deploy have separate hydra_runner and monitor_status — use the deploy versions in deploy scripts. ExperimentConfig comes from nvidia_tao_core in both (same schema/field paths).

Phase 3+4 gate: all three in-container checks pass (tao-pytorch imports + model + ONNX export; tao-deploy imports).

Phase 5 — Packaging & L0 Testing

Goal: register the model as a console_script in both repos and add unit tests.

Steps 12–15: register '<model_name>=...:main' in console_scripts of tao-pytorch/setup.py (12) and tao-deploy/setup.py (13, creating the deploy entrypoint/<model_name>.py via entrypoint_hydra); deploy L0 tests (14); trainer L0 tests — Trainer(..., fast_dev_run=True) + @pytest.mark.cv_unit @pytest.mark.<model_name> (15). See the Phase 5 reference for exact entry-point strings, code, and L0 test file lists.

Gate: entrypoints registered; pytest files exist and follow the marker convention. Do NOT stop — go directly to Phase 6.

Cross-Phase Data Flow & Consistency Verification

Before Docker testing, verify the chain train → export → gen_trt_engine → inference / evaluate (the *_model_latest.pth → .onnx → .engine artifact flow + the config fields each stage reads/writes — full diagram in references/cross-cutting.md).

Consistency checklist (verify before proceeding): self.checkpoint_filename → the *_latest.pth name evaluate.checkpoint / export.checkpoint reference; augmentation.mean/std identical across training spec, inference.yaml, evaluate.yaml, engine-builder preprocess_mode; ONNX input_names=['input'] / output_names=['output'] (detection/instance-seg use task-specific names); export.input_width/input_height match dataset.img_size; model.head.in_channels matches model_params_mapping.py; classes.txt at dataset.root_dir readable by both repos; all __init__.py exist (incl. scripts/__init__.py for get_subtasks() via pkgutil). Full paths: workflow-consistency.md.

Phase 6 — Container Testing & End-to-End Validation

Mandatory — start immediately after Phase 5. All TAO models ship as Docker images; code that only works outside a container is incomplete. Testing runs directly inside the TAO Toolkit container — no image build in the loop: mount → install source (setup.py develop) → run pytest / pylint / pydocstyle / flake8 directly. Use vanilla commands, NOT the ci/run_functional_tests.py / ci/run_static_tests.py wrappers (internal-mirror-only; public github.com/NVIDIA-TAO/ mirrors have no ci/ dir).

Steps 16–25: verify local image tags exist (16); unit tests for tao-core / tao-pytorch (-m cv_unit, --shm-size=16G) / tao-deploy (17–19); lint (20); wheels (21); end-to-end — train dry-run + export in one tao-pytorch session, then gen_trt_engine + inference + evaluate in one tao-deploy session (same session critical — --rm discards installs) (22); cross-check native vs TRT (23); debug shells (24); optional release images (25).

Full commands (every docker run, per-container env-vars, exact pytest / lint invocations + full-suite variants, the train/export/gen_trt_engine/inference/evaluate one-liner with all CLI overrides, the ci/ note, the fix-and-retest loop) and build scripts / runner patterns: see the Phase 6 references.

Phase 6 gate (Done criteria): tao-core / tao-pytorch / tao-deploy unit tests pass in their containers; static tests pass (or only legacy lint warnings); wheels build; end-to-end <model_name>_model_latest.pth → model.onnx → model.engine → non-empty result.csv + results.json; native vs TRT agree within tolerance.

Phase 7 — Optimization & Tuning (conditional)

Enter only if Phase 6 passes but accuracy / latency / size needs improvement. Ask the user for target metrics first.

Diagnostic categories: accuracy too low; TRT-vs-native gap; training too slow; inference too slow. Techniques: Step 27 — hyperparameter tuning; Step 28 — INT8 quantization (PTQ via torchao / modelopt, TRT INT8 + calibration); Step 29 — channel pruning + retrain; Step 30 — knowledge distillation; Step 31 — resolution tuning (TAO interpolates ViT positional embeddings automatically). See the Phase 7 reference for each category's checks, config blocks, YAML overrides, decision tree, and rationale.

Argument

$ARGUMENTS

If provided, interpret $ARGUMENTS as the HuggingFace model ID or URL to start Phase 1. If credentials or model short-name are not included, ask the user for them before proceeding.

Files

references/task-type-guide.md

<!-- Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->

CV Task Type Guide

Use this reference only when the parent SKILL.md points here for the current task. If this file conflicts with current SKILL.md, skill_info.yaml, schemas, or platform/model skills, the current authoritative source wins.

Contents

  • Quick Reference

    1. Image Classification

    • Architecture

    • Implementation Notes

    • Config Specifics

    • Dataset Structure

    • ONNX Export

    • TRT Deploy

    1. Object Detection (DETR-based)

    • Architecture

    • Implementation Notes

    • Config Specifics

    • Dataset Structure

    • Loss Functions

    • ONNX Export

    • TRT Deploy

    • Metrics

    1. Semantic Segmentation

    • Architecture

    • Implementation Notes

    • Config Specifics

    • Dataset Structure

    • Loss Functions

    • ONNX Export

    • TRT Deploy

    • Metrics

    1. Instance Segmentation

    • Architecture

    • Implementation Notes

    • Config Specifics

    • ONNX Export

    • Post-Processing

How to adapt the TAO implementation for each Computer Vision task type. The HF model's pipeline_tag determines which patterns to follow.

Quick Reference

pipeline_tagTAO Reference ModelOutputsONNX OutputsPost-ProcessingMetricsDataset Format
image-classificationclassification_pytLogits (B,C)SingleSoftmax → argmaxTop-k AccuracyClass subdirectories
object-detectiondino, rtdetrLogits (B,Q,C) + Boxes (B,Q,4)Multi (2)Sigmoid → Top-KmAP (COCO)COCO JSON
image-segmentationsegformerLogits (B,C,H,W)SingleArgmax per pixelmIoUImage + Mask pairs
instance-segmentationmask2formerLogits (B,Q,C) + Masks (B,Q,H,W)Multi (2)Threshold + filterAP (COCO)COCO JSON + masks
panoptic-segmentationoneformerLogits + MasksMulti (2)Merge stuff+thingsPQCOCO Panoptic
zero-shot-object-detectiongrounding_dinoLogits (B,Q,T) + Boxes (B,Q,4)Multi (2+text)Contrastive scoremAPCOCO JSON + captions
depth-estimationmono_depthDepth map (B,1,H,W)SingleDirect outputRMSE, Abs.RelImage + depth maps

Key: B=batch, C=classes, Q=queries, H/W=spatial, T=text tokens

1. Image Classification

Architecture
Backbone (ViT, ResNet, etc.) → Global Pooling → Linear Head → Logits (num_classes)
Implementation Notes

  • Simplest task type — use classification_pyt as the direct reference

  • Single output tensor, no post-processing beyond softmax+argmax

  • BackboneBase.get_classifier() returns the linear head

  • BackboneBase.forward() returns logits directly

Config Specifics
ModelConfig:
  backbone: BackboneConfig     # type, pretrained_path, freeze
  head: HeadConfig             # type=TAOLinearClsHead, in_channels, topk, loss
Dataset Structure
root_dir/
├── classes.txt              # Alphabetically sorted class names
├── train/{class_name}/      # Images organized by class
├── val/{class_name}/
└── test/{class_name}/
ONNX Export
input_names=["input"]         # (B, 3, H, W)
output_names=["output"]       # (B, num_classes)
TRT Deploy

  • Reuse ClassificationEngineBuilder, ClassificationInferencer, ClassificationLoader from classification_tf1/

  • preprocess_mode="torch" for ImageNet normalization

2. Object Detection (DETR-based)

Architecture
Backbone → Input Projection (1x1 Conv + GroupNorm per scale level)
  → Deformable Transformer Encoder → Transformer Decoder
  → Class Head (Linear → num_classes) + Box Head (MLP → 4)
Implementation Notes

  • Multi-scale features: backbone produces feature pyramid at strides [4, 8, 16, 32]

  • Use backbone.forward_feature_pyramid(x) instead of backbone.forward(x)

  • Hungarian matching loss (optimal assignment between predictions and GT)

  • DETR models are NMS-free — use Top-K selection instead

  • DN (denoising) queries require special handling during training

  • Detection head outputs normalized box coords (cx, cy, w, h) — convert to (x1, y1, x2, y2) in post-processing

Config Specifics
ModelConfig:
  backbone: str                    # backbone variant name
  num_queries: int = 300           # number of detection queries
  num_feature_levels: int = 4      # multi-scale feature levels
  enc_layers: int = 6              # encoder layers
  dec_layers: int = 6              # decoder layers
  hidden_dim: int = 256            # transformer hidden dim
  cls_loss_coef: float = 2.0       # classification loss weight
  bbox_loss_coef: float = 5.0      # L1 box loss weight
  giou_loss_coef: float = 2.0      # GIoU loss weight
Dataset Structure
data_dir/
├── train/
│   ├── images/
│   └── annotations.json       # COCO format
├── val/
│   ├── images/
│   └── annotations.json
└── classmap.txt               # For inference: class names
Loss Functions

  • Sigmoid Focal Loss: Classification (alpha=0.25, gamma=2)

  • L1 Loss: Box regression on normalized coords

  • GIoU Loss: Generalized IoU for box alignment

  • Hungarian Matching: scipy.optimize.linear_sum_assignment for bipartite matching

  • Auxiliary losses: From intermediate decoder layers

ONNX Export
input_names=["input"]           # (B, 3, H, W)
output_names=["pred_logits", "pred_boxes"]
# pred_logits: (B, num_queries, num_classes) — raw logits, sigmoid in post-processing
# pred_boxes: (B, num_queries, 4) — normalized (cx, cy, w, h)
TRT Deploy

  • Use DDETRDetEngineBuilder (from deformable_detr) — also used by DINO

  • Use DDETRInferencer — handles multi-output extraction

  • Post-processing: sigmoid on logits → Top-K selection → box coord scaling → (x1,y1,x2,y2)

  • Output: annotated images + KITTI-format label files

Metrics

3. Semantic Segmentation

Architecture
Backbone → Multi-scale Feature Pyramid
  → Decode Head (feature fusion + upsampling) → Per-pixel Logits (num_classes, H, W)
Implementation Notes

  • Use backbone.forward_feature_pyramid(x) for multi-scale features

  • Decode head fuses features at multiple resolutions

  • Output spatial dimensions match input (or can be lower-res + bilinear upsample)

  • Loss computed per-pixel with optional ignore index (e.g., 255 for void)

  • SegFormerHead uses multi-resolution MLP fusion

Config Specifics
ModelConfig:
  backbone: BackboneConfig
  decode_head: DecodeHeadConfig
    in_channels: [64, 128, 320, 512]    # Per-scale feature dimensions
    in_index: [0, 1, 2, 3]              # Which backbone scales to use
    feature_strides: [4, 8, 16, 32]     # Spatial stride per scale
    decoder_params:
      embed_dim: 256                     # Decoder hidden dim

DatasetConfig:
  segment:
    palette:                             # Label-to-color mapping
      - {label_id: 0, rgb: [0,0,0], mapping_class: "background", seg_class: "background"}
      - {label_id: 1, rgb: [128,0,0], mapping_class: "person", seg_class: "person"}
    label_transform: "norm"              # or None
Dataset Structure
data_dir/
├── train/
│   ├── images/       # RGB images
│   └── masks/        # Single-channel PNG, pixel value = class_id
├── val/
│   ├── images/
│   └── masks/
└── test/
    ├── images/
    └── masks/
Loss Functions

  • Cross Entropy: Per-pixel classification (supports ignore_index=255)

  • Focal Loss: Hard example mining (alpha, gamma configurable)

  • mIoU Loss: Directly optimizes Intersection over Union

  • mmIoU Loss: Minimax IoU — encourages balanced class performance

ONNX Export
input_names=["input"]           # (B, 3, H, W)
output_names=["output"]         # (B, num_classes, H, W)
# Dynamic spatial dims: H and W can vary
dynamic_axes={"input": {0: "batch", 2: "height", 3: "width"},
              "output": {0: "batch", 2: "height", 3: "width"}}
TRT Deploy

  • Use SegformerEngineBuilder (minimal override of base builder)

  • Use SegformerInferencer + SegformerLoader (from segformer/)

  • Post-processing: argmax per pixel → save as PNG mask

  • Output: mask PNGs + optional overlay visualizations

Metrics

  • mIoU (mean Intersection over Union) — primary

  • Per-class IoU

  • Pixel accuracy

4. Instance Segmentation

Architecture
Backbone → Pixel Decoder (multi-scale feature refinement)
  → Transformer Decoder (query-based instance prediction)
  → Class Head + Mask Head
Implementation Notes

  • Outputs per-instance masks (not per-class like semantic seg)

  • Each query predicts one instance: class + binary mask

  • Mask predictions at reduced resolution — upsampled in post-processing

  • Uses Hungarian matching (like detection) to assign predictions to GT instances

  • Supports both "thing" (countable) and "stuff" (uncountable) categories

Config Specifics
ModelConfig:
  backbone: BackboneConfig
  num_queries: int = 100        # One query per potential instance
  # Mask head and class head integrated into transformer decoder
ONNX Export
input_names=["input"]                    # (B, 3, H, W)
output_names=["pred_logits", "pred_masks"]
# pred_logits: (B, num_queries, num_classes + 1)  — includes no-object class
# pred_masks: (B, num_queries, H/4, W/4)          — reduced resolution
Post-Processing

  1. Softmax on logits → filter by confidence threshold

  2. Select Top-K instances by score

  3. Bilinear upsample masks to original resolution

  4. Apply sigmoid → threshold at 0.5 for binary masks

  5. Remove overlapping instances (higher confidence wins)

Metrics

5. Panoptic Segmentation

Architecture

Same as instance segmentation, but with task-conditional head that handles both "things" and "stuff":

Backbone → Pixel Decoder → Transformer Decoder
  → Task-conditional Head (semantic + instance + panoptic modes)
Implementation Notes

  • Unified architecture for semantic, instance, and panoptic segmentation

  • Task token conditions the decoder behavior

  • Stuff classes (background, sky): treated like semantic seg

  • Thing classes (person, car): treated like instance seg

  • Panoptic output merges both

Metrics

  • PQ (Panoptic Quality) = SQ × RQ

    • SQ (Segmentation Quality): IoU of matched segments

    • RQ (Recognition Quality): F1 of matched/unmatched

  • Also reports mIoU for stuff and AP for things

6. Zero-Shot / Grounding Detection

Architecture
Image Backbone → Multi-scale Features
Text Encoder (BERT) → Text Embeddings
  → Cross-Modal Fusion (Transformer)
  → Contrastive Class Head + Box Head
Implementation Notes

  • Requires text input in addition to images — major architectural difference

  • Class predictions via contrastive similarity (not fixed linear head)

  • Text encoder (BERT) can be frozen or fine-tuned

  • Feature alignment layer maps text embeddings to vision space

  • ONNX export must handle text input tensors

  • At inference, text prompt defines what to detect (open vocabulary)

ONNX Export
input_names=["inputs", "input_ids", "attention_mask", "position_ids",
             "token_type_ids", "text_token_mask"]
output_names=["pred_logits", "pred_boxes"]
# pred_logits shape: (B, num_queries, max_text_len) — NOT num_classes!
Special Considerations

  • Text tokenizer needed at inference time (pre-tokenize or include in pipeline)

  • Logit shape depends on text length, not fixed class count

  • Contrastive scoring: aggregate logits across text tokens per detection

7. Depth Estimation

Architecture
Encoder (Backbone) → Decoder (progressive upsampling) → Depth Map (1, H, W)
Implementation Notes

  • Single-channel output (depth value per pixel)

  • May use photometric loss (compares reprojected views)

  • Stereo variants take two input images

ONNX Export
input_names=["input"]           # (B, 3, H, W) or (B, 6, H, W) for stereo
output_names=["output"]         # (B, 1, H, W)
Metrics

  • RMSE (Root Mean Square Error)

  • Abs.Rel (Absolute Relative Error)

  • δ thresholds (% of pixels within 1.25^n ratio)

Task-Type Decision Tree

When the agent determines the HF model's pipeline_tag, use this to select the implementation strategy:

pipeline_tag
├── image-classification
│   └── Reference: classification_pyt
│       └── Simple backbone + linear head
│       └── Single ONNX output
│       └── Reuse Classification{EngineBuilder,Inferencer,Loader}
│
├── object-detection
│   └── Reference: dino (DETR-based) or rtdetr (real-time)
│       └── Backbone + transformer encoder/decoder + detection heads
│       └── Multi ONNX output (logits + boxes)
│       └── Hungarian matching loss
│       └── Needs DDETRDet{EngineBuilder,Inferencer}
│
├── image-segmentation
│   └── Reference: segformer
│       └── Backbone + decode head
│       └── Single ONNX output (spatial)
│       └── Per-pixel loss with ignore_index
│       └── Reuse or extend Segformer{EngineBuilder,Inferencer,Loader}
│
├── instance-segmentation
│   └── Reference: mask2former
│       └── Backbone + pixel decoder + transformer decoder
│       └── Multi ONNX output (logits + masks)
│       └── Hungarian matching + mask losses
│
├── panoptic-segmentation
│   └── Reference: oneformer
│       └── Task-conditional architecture
│       └── Multi ONNX output + task token
│
├── zero-shot-object-detection
│   └── Reference: grounding_dino
│       └── Multi-modal (image + text)
│       └── BERT text encoder required
│       └── Contrastive class prediction
│
├── depth-estimation
│   └── Reference: mono_depth / stereo_depth
│       └── Encoder-decoder for depth maps
│       └── Single ONNX output
│
└── OTHER
    └── Halt — unsupported task type

What Changes Per Task Type

ComponentClassificationDetectionSegmentationInstance Seg
backbone.forwardforward()forward_feature_pyramid()forward_feature_pyramid()forward_feature_pyramid()
Head typeLinearTransformer + MLPDecode headPixel decoder + Transformer
LossCEFocal + L1 + GIoUCE / Focal / IoUCE + Mask + Match
ONNX outputs12 (logits, boxes)12 (logits, masks)
Dynamic spatialNoYes (image H/W)Yes (H/W)Yes (H/W)
Post-processSoftmax+argmaxSigmoid+TopK+scaleArgmax per pixelSigmoid+filter+upsample
DatasetClass dirsCOCO JSONImage+Mask pairsCOCO JSON + masks
Deploy inferencerClassification*DDETR*Segformer*Custom
Deploy dataloaderClassification*ImageBatcherSegformer*Custom
MetricsTop-k AccmAP (COCO)mIoUAP (COCO)

*Reusable from existing TAO implementations

Positional Embedding Handling for ViT-based Models

When the HF model uses a ViT backbone and the TAO training resolution differs from the pretrained resolution:

# TAO handles this automatically in backbone_v2/vit.py:
def _interpolate_pos_encoding(self, x, w, h):
    # Bicubic interpolation of positional embeddings
    # Class token kept unchanged, patch tokens interpolated
    pos_tokens = F.interpolate(pos_tokens, size=(new_h, new_w), mode='bicubic')

Also available as a utility:

from nvidia_tao_pytorch.core.utils.pos_embed_interpolation import interpolate_pos_embed
checkpoint = interpolate_pos_embed(checkpoint, orig_resolution, orig_patch_size,
                                    new_resolution, new_patch_size)

This is critical when the HF model was pretrained at e.g., 224×224 but TAO trains at 384×384.

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 Python, Needs Node.js

The author says: Requires Python 3.10+, NVIDIA driver, CUDA 13.0+, docker + nvidia-container-toolkit, an NGC API key (`docker login nvcr.io`), an HF_TOKEN, and access to the TAO Toolkit container images on `nvcr.io` for `tao-pytorch`, `tao-deploy`, and (optionally) `tao-dataservices` — Phase 0 asks the user for the exact image references and prepares them locally as `tao-pytorch-base:latest`, `tao-deploy-base:latest`, `tao-dataservices-base:latest`. Local clones of `tao-core`, `tao-pytorch`, `tao-deploy`,...

Security

What our scanner noticed· Medium risk

  • Large base64-encoded blobs
  • Instructions that use the network

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-port-huggingface-model · commit fce3cca
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE-APACHE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

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-port-huggingface-model.zip