Tao Finetune Clip

CLIP

Standalone install? Read CLIP standalone setup and verify the bank checkout before running tao-setup or launching a CLIP action.

Contrastive Language-Image Pre-training model for zero-shot and fine-tuned image classification, image-text retrieval, and embedding extraction. Fine-tuning adapts CLIP's shared image-text embedding space to domain-specific image-caption data.

No default NGC pretrained checkpoint is required for spec construction, but unset checkpoint behavior is action-specific. In the validation-fixes PyTorch image, export.checkpoint: null exports the selected CLIP architecture and may initialize weights when pretrained weights are unavailable. Do not assume inference.checkpoint: null loads pretrained weights: clip inference currently calls the checkpoint loader with None and fails before embedding extraction. For PyTorch inference, checkpoint-backed evaluation/export, resume, and retrain flows, resolve and pass an exact checkpoint from the parent train output. For trusted TAO checkpoints produced by the current run or a known parent job, set TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 on checkpoint-dependent PyTorch actions so PyTorch 2.6 can load the Lightning checkpoint metadata; do not set this for untrusted checkpoints.

Supported actions: train, evaluate, inference, export, gen_trt_engine.

Train Action Policy

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read references/skill_info.yaml and resolve the run override from either an explicit automl_policy value or the user's workflow request. Use automl_policy: on by default and only expose on / off in new launch prompts. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as automl_policy: off for this run only. When automl_policy: on, automl_enabled: true, and both schemas/train.schema.json and references/spec_template_train.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skill_dir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automl_policy. Use direct model training only when automl_policy: off or the packaged train schema/template is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.

The packaged CLIP train schema enables train.optim.vision_lr and train.optim.text_lr as default AutoML search parameters. For smoke tests, keep the search small by using the Bayesian algorithm with two recommendations and narrow LR ranges.

PAS fine-tuning method integration

For every new PAS DEFT run, ask one mutually exclusive question when the user has not already chosen a method:

Fine-tuning method: LoRA (default) or full-parameter SFT?

This LoRA-default contract is scoped to PAS. Standalone CLIP training keeps its packaged SFT image and template behavior; it does not default to LoRA. Do not transfer PAS image capabilities or this selection table to standalone CLIP.

An image capability probe is a launch-gated docker run, so include it in the approval summary before executing it. Documentation, an image tag, or an unrecognized Hydra override is not proof of LoRA support. If the probe does not expose a complete adapter contract (enable flag/config, target modules, rank, alpha, dropout, checkpoint behavior), report LoRA as unsupported for that image and stop before materializing a run.

Non-train actions such as evaluate, inference, export, and deploy flows stay in this model skill. The per-run automl_policy override does not change model metadata.

Instructions

Use this skill for NVIDIA TAO CLIP jobs: training, evaluation, embedding inference, ONNX export, and TensorRT engine generation. Start by identifying the requested action, then read references/skill_info.yaml for image, command, and data-source wiring; references/spec_template_<action>.yaml for packaged action defaults; and schemas/<action>.schema.json when packaged for parameter metadata. Use references/spec_template.yaml for the full spec shape and references/tao-deploy-clip.md for TensorRT actions.

For dataset-backed actions, collect the required image, caption, list, or prompt files from the user and place the resolved paths in spec_overrides. For local Docker runs, mount extracted folders in the container and point image_dir / caption_dir at those folders; if a data source provides .tar.gz archives, extract them before running the in-container CLIP commands. For export and gen_trt_engine, infer parent artifacts from the upstream job when available; otherwise require explicit checkpoint, ONNX, or engine paths. Run gen_trt_engine, TensorRT evaluate, and TensorRT inference in the TAO Deploy image.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-clip.md first. Deploy spec templates live in this skill's references/ folder with the spec_template_deploy_*.yaml prefix.

Training Requirements

• Dataset type: image_text

• Formats: custom image/caption folders or WebDataset shards

• Monitoring metric: val/t2i_mAP for paired-caption retrieval; val/pas/overall_mAP for PAS metadata matching

Before selecting a PAS AutoML objective or comparing training and evaluation, read PAS metric guidance. PAS validation logs val/pas/{easy,medium,hard,overall}_{mAP,rank1,rank5}; standalone PAS evaluate writes CSVs rather than test/* scalars. For a like-for-like comparison, use the same checkpoint and validation subset, set evaluate.pas_ground_truth_mode: scalar_attributes, and read nvidia_pas_metadata_metrics_weighted_aggregate.csv.

Supported Models

• OpenCLIP / NV-CLIP: ViT-L-14-SigLIP-CLIPA-224 (default), ViT-L-14-SigLIP-CLIPA-336, ViT-H-14-SigLIP-CLIPA-224, ViT-H-14-SigLIP-CLIPA-336, ViT-H-14-SigLIP-CLIPA-574

• Radio-CLIP: c-radio_v3-b, c-radio_v3-l, c-radio_v3-h, c-radio_v3-g

• SigLIP2: siglip2-so400m-patch16-256, siglip2-so400m-patch14-224, siglip2-so400m-patch14-384, siglip2-so400m-patch16-384, siglip2-so400m-patch16-512, siglip2-so400m-patch16-naflex

Radio-CLIP requires model.adaptor_name to be set to siglip or clip.

Under the tower mode contract, OpenCLIP supports full / frozen, but rejects lora. See backbone support.

LoRA and preservation regularization

Before constructing a LoRA spec or PEFT AutoML search space, read CLIP PEFT guidance for per-tower fields, defaults, backbone target modules, numeric bounds, migration from tower enabled to mode, and encoder-freeze precedence. peft and regularization are top-level spec blocks; use the selected runtime's config contract.

With PEFT enabled under the tower mode contract, tower modes override model.freeze_*_encoder: frozen freezes the tower, full trains it, and lora trains adapters with the backbone frozen. Remove conflicting freeze flags and verify per-tower trainable-parameter counts.

LoRA checkpoints contain the full model; budget full-model storage. Preservation regularization adds a full frozen teacher on the GPU and a teacher forward pass. Keep regularization.enabled: false for recommended SigLIP2 PAS runs: preservation loss has not been fully validated there. The teacher is excluded from checkpoints.

Per-Action Dataset Requirements

For custom training, set dataset.train.type: custom and provide dataset.train.datasets entries. Image and caption files must share the same base name. caption_file_suffix defaults to .txt, and image_list_file is optional.

To run inference on a subset, place only the selected images in a separate directory and set inference.datasets[].image_dir to that directory.

When no native CLIP image-caption dataset is available, do not silently treat image-classification data as CLIP data. If the user explicitly allows a plumbing-only validation fallback, derive caption files from class labels, document that the captions are generated from labels, and keep each image/caption pair on the same base filename. Without an image_list_file, the TAO custom loader scans the configured image directory for image files; keep validation folders flat unless you provide a list file.

For WDS training, set dataset.train.type: wds and provide at least one of dataset.train.wds.root_dir or dataset.train.wds.shard_list_file. root_dir is scanned recursively for .tar shards. shard_list_file is a text file with one shard path per line; relative lines resolve under the list-file directory unless root_dir is also supplied, in which case they resolve under root_dir. Validation/evaluation data remains custom format via dataset.val.datasets.

Typical Spec Overrides

Data source overrides are mandatory for dataset-backed actions. Construct paths from the Per-Action Dataset Requirements table and include them in spec_overrides. For inference, provide at least one of inference.datasets or inference.text_file.

S3_TRAIN = "s3://bucket/data/train"
S3_WDS = "s3://bucket/data/wds"
S3_EVAL = "s3://bucket/data/eval"
S3_INFER = "s3://bucket/data/infer"

train, custom dataset:

{
    "train.num_epochs": 10,
    "dataset.train.type": "custom",
    "dataset.train.datasets": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "image_list_file": f"{S3_TRAIN}/image_list.txt", "caption_dir": f"{S3_TRAIN}/captions.tar.gz"}],
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

train, WDS dataset:

{
    "train.num_epochs": 10,
    "dataset.train.type": "wds",
    "dataset.train.wds.root_dir": f"{S3_WDS}",
    "dataset.train.wds.shard_list_file": f"{S3_WDS}/shards.txt",
    "dataset.train.wds.samples_per_shard": 10000,
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

evaluate:

{
    "dataset.val.datasets": [{"image_dir": f"{S3_EVAL}/images.tar.gz", "image_list_file": f"{S3_EVAL}/image_list.txt", "caption_dir": f"{S3_EVAL}/captions.tar.gz"}],
}

Leave evaluate.checkpoint unset for zero-shot evaluation with pretrained weights. Set evaluate.trt_engine instead of evaluate.checkpoint for TensorRT evaluation.

inference:

{
    "inference.datasets": [{"image_dir": f"{S3_INFER}/images.tar.gz"}],
    "inference.text_file": f"{S3_INFER}/prompts.txt",
}

Inference writes image_embeddings.h5 and/or text_embeddings.h5 under results_dir. The saved embeddings are L2-normalized.

export:

{
    "export.onnx_file": "${results_dir}/export/clip_model.onnx",
    "export.encoder_type": "combined",
    "export.batch_size": -1,
}

Set export.encoder_type: separate when deployment should use independent vision and text encoders. Separate export writes _vision.onnx and _text.onnx variants derived from the base export.onnx_file.

For checkpoint-dependent actions, use the model-specific checkpoint resolver output from the parent train job. CLIP training writes checkpoints such as model_epoch_000_step_00020.pth and a clip_latest.pth symlink. Use the exact resolved checkpoint for evaluate.checkpoint, inference.checkpoint, export.checkpoint, and train.resume_training_checkpoint_path; use clip_latest.pth only when the user explicitly asks for latest.

For LoRA checkpoints, preserve the training model settings and complete top-level peft block in PyTorch evaluate, inference, and export specs. Read checkpoint configuration before constructing those specs; action defaults alone omit PEFT.

When the resolved checkpoint is trusted TAO output, checkpoint-backed PyTorch evaluate, inference, export, and resume training should run with TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1. PyTorch 2.6 otherwise defaults checkpoint loading to weights-only mode and can reject CLIP Lightning checkpoints containing NumPy scalar metadata.

gen_trt_engine:

{
    "gen_trt_engine.onnx_file": "${results_dir}/export/clip_model.onnx",
    "gen_trt_engine.trt_engine": "${results_dir}/deploy/clip_model.engine",
    "gen_trt_engine.batch_size": -1,
    "gen_trt_engine.tensorrt.data_type": "fp16",
    "gen_trt_engine.tensorrt.min_batch_size": 1,
    "gen_trt_engine.tensorrt.opt_batch_size": 1,
    "gen_trt_engine.tensorrt.max_batch_size": 16,
}

Eval Dataset

Optional for training. If provided, validation metrics are computed at validation intervals. Required for evaluate.

Deploy Workflow

The skill exposes gen_trt_engine as the deploy action. In generated SDK runners, use model_info["actions"]["gen_trt_engine"] and run it in the TAO Deploy image, not the PyTorch training image. The in-container command is clip gen_trt_engine -e {config_path}; direct TAO Launcher usage spells the same action as tao deploy clip gen_trt_engine -e /path/to/spec.yaml.

TAO Deploy inference can discover combined engines, paired separate engines, or single-pillar _vision.engine / _text.engine files. For full TensorRT retrieval evaluation or image+text TensorRT inference, export with export.encoder_type: separate and run clip gen_trt_engine twice: build clip_model_vision.onnx to an engine ending in _vision.engine, then build clip_model_text.onnx to the matching _text.engine in the same directory. For image-only TensorRT inference, building only the _vision.engine is sufficient and inference.text_file must be null. TensorRT evaluate and text inference require a text-capable engine; if only a vision engine is present, deploy evaluation fails because text embeddings cannot be extracted.

Use evaluate.trt_engine for TensorRT evaluation and inference.trt_engine for TensorRT embedding extraction. These TensorRT paths also run in the TAO Deploy image. Direct TAO Launcher usage spells these as tao deploy clip evaluate and tao deploy clip inference.

Important Parameters

• model.type: Backbone family and resolution. Use a TAO-registered CLIP model ID such as ViT-L-14-SigLIP-CLIPA-224. Prefer the listed OpenCLIP / NV-CLIP IDs for AutoML smoke tests because the current TAO container registry routes them through the supported augmentation adapter.

• model.adaptor_name: Required for Radio-CLIP. Set to siglip or clip.

• model.image_size: Training transform image resolution. Keep it aligned with the selected fixed-resolution backbone.

• train.num_epochs: CLIP fine-tuning often converges quickly. Start with 10-20 epochs for domain adaptation, then increase only if validation loss is still improving.

• train.optim.vision_lr / train.optim.text_lr: Learning rates for the two encoders. CLIP is sensitive to high learning rates; reduce both if loss is unstable.

• model.freeze_vision_encoder / model.freeze_text_encoder: Defaults are false. These flags control encoder freezing when peft.enabled: false. With PEFT enabled under the tower mode contract, peft.vision.mode / peft.text.mode override them, even if they disagree: full trains the encoder, frozen freezes it, and lora trains adapters. See PEFT precedence and remove conflicting freeze flags.

• train.loss_type: siglip is recommended for SigLIP2 and Radio-CLIP. Use clip for CLIP-style softmax loss.

• export.encoder_type: combined exports one ONNX graph. separate exports independent vision and text graphs.

• gen_trt_engine.tensorrt.data_type: TensorRT deployment supports fp16 and fp32.

Hardware

Single-GPU training works for small datasets. Use 4+ GPUs for datasets with more than 100k images or large backbones. Use 16GB+ VRAM per GPU for small/fixed-resolution runs with preservation regularization disabled and larger GPUs for Radio-CLIP or high-resolution OpenCLIP variants.

The 16GB+ guidance does not account for the additional full teacher model created by regularization.enabled: true. If testing regularization, budget VRAM for that second model and its forward-pass intermediates, measure peak memory and step time for the actual backbone and batch size, and reduce dataset.train.batch_size or use a larger GPU as needed. Extra memory depends on the model's resident dtype; train.precision: fp16 alone does not guarantee that the teacher weights occupy half precision.

Error Patterns

See references/error-patterns.md for the full list of CLIP error symptoms and fixes (CUDA OOM, NaN loss, retrieval quality, dataset format/size, Radio-CLIP and model-ID validation, ONNX external data, TensorRT shape mismatch, PyTorch 2.6 checkpoint load, null-checkpoint inference, TensorRT text/retrieval failures, attention_mask handling, and spec/schema merge errors).

Spec Param / Parent Model Inference

See references/spec-param-inference.md for the packaged model-specific action/spec-field/inference-function mappings that generated runners apply with SDK helpers before create_job(), plus parent_job_id resolution rules.

Deployment

• tao-deploy-clip