nvidia-tao/tao-skill-bank/tao-finetune-video-clip

InternVideo2-CLIP (TAO video_clip)

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).

TAO task `video_clip` wraps OpenGVLab InternVideo2-CLIP L14. The PyTorch image provides train, evaluate, inference, export, and default_specs. TAO Deploy provides gen_trt_engine, TensorRT evaluate, and TensorRT inference.

Container images and per-action commands are in references/skill_info.yaml and references/tao-deploy-video-clip.skill_info.yaml. Starting specs are in references/spec_template_*.yaml.

Release note: The pinned PyTorch image is the TAO 7.2 release-candidate build validated for Video-CLIP. It includes PyAV 17.1.0 as the primary decoder and ONNXScript 0.7.1 for export, with decord absent. The TAO Deploy image is pinned independently because gen_trt_engine and TensorRT-backed actions do not run in the PyTorch image.

Known-broken images: interim builds cut before tao-pytorch commit 0cc31de4 ship a video_clip package with no model.backbones submodule, so train/evaluate/inference die at import while video_clip --help still exits 0. Images without PyAV also fail at data loading. Run both import checks in the preflight below before pulling data or launching a run.

Train Action Policy

AutoML is not packaged for this model skill. Always use direct video_clip actions even when a higher-level request mentions AutoML. Non-train actions stay in this skill.

Quick Start (local Docker)

Use the pinned TAO container declared in references/skill_info.yaml. Pull with NGC_KEY when the image is not cached locally.

VIDEO_CLIP_IMAGE_DEFAULT="nvcr.io/nvidia/tao/tao-toolkit:7.3.0-pyt"  # versions-key: images.tao_toolkit.pyt
VIDEO_CLIP_IMAGE="${VIDEO_CLIP_IMAGE:-$VIDEO_CLIP_IMAGE_DEFAULT}"
docker pull "$VIDEO_CLIP_IMAGE"

Expected workspace layout (host paths bind-mounted into the container):

workspace/
├── data/
│   ├── train.json              # vadr1_chunks metadata; video_path = /data/videos/<name>.mp4
│   ├── val.json
│   ├── prompts.txt             # one text prompt per line (inference)
│   └── videos/                 # mp4 clips referenced by train.json / val.json
├── model/
│   ├── mobileclip_blt.pt       # MobileCLIP weights for model.text_encoder
│   └── hf/                     # offline InternVideo2_distillation_models snapshot (HF_HUB_OFFLINE=1)
├── specs/
│   ├── train.yaml
│   ├── evaluate.yaml
│   ├── inference.yaml
│   └── export.yaml
├── deploy_specs/
│   ├── gen_trt_engine.yaml
│   ├── evaluate.yaml
│   └── inference.yaml
└── results/

Docker options for all actions (skill-eval CI uses the same $WORKSPACE_DIR bind-mount pattern):

VIDEO_CLIP_IMAGE_DEFAULT="nvcr.io/nvidia/tao/tao-toolkit:7.3.0-pyt"  # versions-key: images.tao_toolkit.pyt
VIDEO_CLIP_IMAGE="${VIDEO_CLIP_IMAGE:-$VIDEO_CLIP_IMAGE_DEFAULT}"
RUN_ROOT="${RUN_ROOT:-$PWD}"
DOCKER_COMMON=(
  --rm --gpus all --ipc=host --network=host
  --shm-size=64g
  --ulimit memlock=-1
  --ulimit stack=67108864
  -e WANDB_DISABLED=true
  -e WANDB_MODE=disabled
  -e HF_HUB_OFFLINE=1
  -e HUGGINGFACE_HUB_CACHE=/model/hf
  -e TRANSFORMERS_OFFLINE=1
  -e TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1
  -v "$RUN_ROOT/data:/data:ro"
  -v "$RUN_ROOT/model:/model:ro"
  -v "$RUN_ROOT/specs:/specs:ro"
  -v "$RUN_ROOT/results:/results"
)

Preflight (host):

[ -f "$RUN_ROOT/model/mobileclip_blt.pt" ] || echo "MISSING: MobileCLIP weights"
[ -f "$RUN_ROOT/data/train.json" ] || echo "MISSING: train metadata"
[ -d "$RUN_ROOT/model/hf" ] || echo "MISSING: offline HF snapshot under model/hf"
docker run --rm "$VIDEO_CLIP_IMAGE" video_clip --help >/dev/null || echo "MISSING: video_clip in container"
docker run --rm "$VIDEO_CLIP_IMAGE" \
  python -c "import nvidia_tao_pytorch.multimodal.video_clip.model.adapters.internvideo2clip" \
  >/dev/null 2>&1 || echo "BROKEN IMAGE: video_clip package is incomplete (missing model.backbones) — stop, see Release note"
docker run --rm "$VIDEO_CLIP_IMAGE" python -c "import av; print(av.__version__)" \
  >/dev/null 2>&1 || echo "BROKEN IMAGE: PyAV is missing — use the pinned FC image; do not install decord"
nvidia-smi >/dev/null 2>&1 || echo "note: no GPU visible"

Train:

docker run "${DOCKER_COMMON[@]}" "$VIDEO_CLIP_IMAGE" \
  video_clip train -e /specs/train.yaml results_dir=/results

Evaluate:

docker run "${DOCKER_COMMON[@]}" "$VIDEO_CLIP_IMAGE" \
  video_clip evaluate -e /specs/evaluate.yaml results_dir=/results

Inference:

docker run "${DOCKER_COMMON[@]}" "$VIDEO_CLIP_IMAGE" \
  video_clip inference -e /specs/inference.yaml results_dir=/results

Export:

docker run "${DOCKER_COMMON[@]}" "$VIDEO_CLIP_IMAGE" \
  video_clip export -e /specs/export.yaml results_dir=/results

TensorRT Deploy

Use the independently pinned TAO Deploy image after PyTorch export. Read references/tao-deploy-video-clip.md before running the deploy actions; its templates cover the engine build, retrieval evaluation, and embedding inference contracts.

VIDEO_CLIP_DEPLOY_IMAGE_DEFAULT="nvcr.io/nvidia/tao/tao-toolkit:7.3.0-deploy"  # versions-key: images.tao_toolkit.deploy
VIDEO_CLIP_DEPLOY_IMAGE="${VIDEO_CLIP_DEPLOY_IMAGE:-$VIDEO_CLIP_DEPLOY_IMAGE_DEFAULT}"
 
# Verify the independently pinned image before staging artifacts or using a GPU.
docker run --rm "$VIDEO_CLIP_DEPLOY_IMAGE" video_clip gen_trt_engine --help >/dev/null || \
  { echo "BROKEN IMAGE: Video-CLIP deploy entrypoint is unavailable" >&2; exit 1; }
 
docker run --gpus all --rm --shm-size=16g \
  -v "$RUN_ROOT/deploy_specs:/specs:ro" \
  -v "$RUN_ROOT/results/export:/models:ro" \
  -v "$RUN_ROOT/data:/data:ro" \
  -v "$RUN_ROOT/results/deploy:/results" \
  "$VIDEO_CLIP_DEPLOY_IMAGE" \
  video_clip gen_trt_engine -e /specs/gen_trt_engine.yaml

Keep the exported ONNX file, its matching *_config.yaml, and its matching *_tokenizer/ directory together. gen_trt_engine copies the sidecars beside the engine so TensorRT evaluate and inference can reconstruct preprocessing and tokenization.

Quick Start (virtualenv — local dev hosts)

On hosts with a tao-pytorch checkout and tao-cli venv (for example rtdetr-pytorch), run through `tao-run-on-virtualenv` instead of Docker:

export VENV="${VENV:?set to your tao-cli virtualenv (must contain bin/video_clip)}"
export TAO_PYTORCH_ROOT="${TAO_PYTORCH_ROOT:?set to your tao-pytorch checkout with multimodal/video_clip}"
export PATH="$VENV/bin:$PATH"
export PYTHONPATH="$TAO_PYTORCH_ROOT:$TAO_PYTORCH_ROOT/tao-core:$PYTHONPATH"
export HF_HOME="${HF_HOME:-$PWD/hf_cache}"
export HF_HUB_OFFLINE="${HF_HUB_OFFLINE:-1}"
export WANDB_DISABLED=true
export WANDB_MODE=disabled

Copy a template spec from references/spec_template_*.yaml, fill checkpoint paths and metadata, then run:

video_clip train -e /path/to/train.yaml
video_clip evaluate -e /path/to/evaluate.yaml
video_clip inference -e /path/to/inference.yaml
video_clip export -e /path/to/export.yaml

Credentials

• NGC_KEY — pull the pinned TAO container from nvcr.io when it is not cached locally.
• HF_TOKEN (online Hugging Face download only): Hugging Face read token used

when model.vision_encoder and model.clip_head are null and the resolver downloads the InternVideo2 snapshot named by model.internvideo2clip_hf_id. For offline CI/eval, stage the complete S3/local snapshot and point both fields at the local files; no Hugging Face token is then required.

Treat tokens as secrets. Export them into the environment or pass them through an --env-file of bare KEY=value lines, rather than inlining values into generated spec YAML, command lines, or anything written under results_dir.

Data format (vadr1_chunks)

Metadata is a top-level JSON list of video records. Each record has video_path, split, and nested chunks[] with caption fields (queries, action_queries, anomaly_queries, dense_caption, scene_caption).

Point dataset.*.video_text.metadata at the user's JSON files (for example /data/train.json and /data/val.json in the templates). Each video_path must be an absolute path resolvable inside the runtime — for Docker, remap host clips to container paths like /data/videos/<name>.mp4 and set data_root: null unless using path_prefix_mapping.

Smoke overrides

For a short functional check (for example 2 epochs, 1 GPU, small batch):

train.num_epochs: 2
train.num_gpus: 1
train.gpu_ids: [0]
train.optim.warmup_steps: 10
dataset.train.batch_size: 2
dataset.val.batch_size: 2
dataset.train.num_workers: 4
dataset.val.num_workers: 4
dataset.train.video_text.caption_fields: [queries, action_queries, anomaly_queries]
dataset.train.video_text.caption_mode: first
dataset.metrics.mode: classification

Use dataset.metrics.mode: retrieval only when dataset.val.video_text.relevance_file is provided.

Inference

• inference.mode: embeddings writes video_embeddings.h5 and text_embeddings.h5 under results_dir.
• Provide inference.query.text_file (one prompt per line) and/or inference.query.input_texts.
• Gallery videos come from dataset.inference.video_text.metadata.

Export

export.encoder_type: combined produces the image-and-text ONNX consumed by the Video-CLIP deploy workflow. Keep export.batch_size: -1 for symbolic/dynamic batch dimensions; a positive value produces a fixed-batch ONNX. Export also writes matching *_config.yaml and *_tokenizer/ sidecars; preserve all three artifacts. Default opset is 23 on the vendor branch. Export requires a trained .pth at export.checkpoint.

LoRA

For vision-LoRA runs, start from tao-pytorch experiment_spec_lora.yaml or add a top-level peft: block (see shipped spec comments). Merge LoRA before export when checkpoints contain lora_* keys.

Common pitfalls