tao-finetune-kumo-anomaly

name: tao-finetune-kumo-anomalydescription: >-  Kumo-Anomaly — diffusion-based anomaly detection and fine-tuning  for multivariate time series. Use when the user asks to "fine-tune Kumo-Anomaly",  "run Kumo-Anomaly inference", "detect anomalies with diffusion", "time series  anomaly detection", "finetune Kumo-Anomaly", "use perform_anomaly_analysis_with_diffusion",  "automl Kumo-Anomaly", "hyperparameter search Kumo-Anomaly", "hyperparameter optimization"  or mentions "curriculum_medium.yaml", "final_model.pth",  "kumo-anomaly", "sdk/anomaly_analysis", or "TSDiffuser_Generic".license: Apache-2.0compatibility: Requires Python 3.12+ and uv. CUDA GPU recommended; CPU-only supported.metadata:   author: NVIDIA Corporation  version: "0.2.0"allowed-tools: Read Bashtags:   - anomaly-detection  - time-series  - diffusion  - finetune  - inference  - automl  - kumo-anomaly

Kumo-Anomaly

Diffusion-based anomaly detection and fine-tuning for multivariate time series. The model reconstructs randomly masked segments and scores each timestep by MAE between reconstruction and original signal; adaptive thresholding (SCS or MACS) converts scores to binary labels.

Source code: https://github.com/NVIDIA/Kumo-TS Pretrained weights: https://huggingface.co/nvidia/Kumo-Anomaly

For the most up-to-date usage information, refer to the README files in the Kumo-TS repository:

• Kumo-Anomaly/README.md — full SDK reference, model architecture, and API docs
• Kumo-Anomaly/examples/datasets/README.md — dataset format, synthetic data generation, and CSV conventions

External dependencies

Credentials

nvidia/Kumo-Anomaly is a public repo — no token required for downloading weights. If you ever hit a 401/403 (gated access or private fork) or a 504 on first download, see the Known pitfalls section.

Quick start

git clone --branch main --single-branch https://github.com/NVIDIA/Kumo-TS Kumo-TS
cd Kumo-TS/Kumo-Anomaly
uv sync                              # install dependencies (one-time)

# Inference — synthetic data, auto-downloads weights from HF on first run
uv run python examples/quick_example.py

# Inference — your own CSV
uv run python examples/quick_example.py \
  --model-path final_model.pth \
  --config-path curriculum_medium.yaml \
  --dataset-path /path/to/data.csv

# Pre-download weights only (warm the cache before going offline)
uv run python examples/quick_example.py --download-weights

# Fine-tune on your own normal-behavior data
uv run python examples/finetune_example.py \
  --csv /path/to/normal_training_data.csv \
  --timestamp-col timestamp \
  --label-col is_anomaly \
  --epochs 20 \
  --output-dir artifacts/finetune_my_data

Inference

Use perform_anomaly_analysis_with_diffusion in sdk/anomaly_analysis.py. It validates input, auto-dispatches across all visible GPUs, applies adaptive thresholding (SCS or MACS), and returns the original DataFrame with Anomaly (0/1) and MAE columns appended.

import sys, pandas as pd
sys.path.append("/path/to/Kumo-TS/Kumo-Anomaly")  # clone Kumo-TS with --branch main
from sdk.anomaly_analysis import perform_anomaly_analysis_with_diffusion

df = pd.read_csv("your_data.csv")

# The API raises ValueError on non-numeric columns — drop timestamp, IDs, and labels first.
df = df.select_dtypes(include="number")

results = perform_anomaly_analysis_with_diffusion(
    df=df,
    threshold_strategy="scs",       # "scs" (fast) or "macs" (adaptive)
    model_path=None,                 # None → auto-download final_model.pth from HF
    model_config_path=None,          # None → auto-download curriculum_medium.yaml from HF
    nsample=15,                      # diffusion samples per window; ↑ accuracy, ↑ latency
    preprocess_model_dir=None,       # optional preprocessing model directory
)
# results columns: Anomaly (0/1), MAE (float), plus all original columns
print(results[["Anomaly", "MAE"]].describe())

Reporting (sdk_config)

perform_anomaly_analysis_with_diffusion takes an optional sdk_config keyword — an ADDiffusionConfig, a path to a YAML file, or None for defaults. It scopes report-generation knobs only; inference-behavior args (threshold_strategy, nsample, model_path, model_config_path) stay as direct function arguments.

from sdk.anomaly_analysis import ADDiffusionConfig, perform_anomaly_analysis_with_diffusion

results = perform_anomaly_analysis_with_diffusion(
    df=df,
    threshold_strategy="scs",
    sdk_config=ADDiffusionConfig(
        report_path="artifacts/anomaly_report.pdf",
        report_title="Anomaly Detection Report",
    ),
)

Or pass a YAML path instead of the dataclass — load_sdk_config accepts a flat mapping or one nested under a top-level sdk: key (so the file can carry other sections alongside it), and rejects unknown keys:

sdk:
  report_path: artifacts/anomaly_report.pdf
  report_title: Anomaly Detection Report

results = perform_anomaly_analysis_with_diffusion(
    df=df,
    threshold_strategy="scs",
    sdk_config="sdk_config.yaml",
)

ADDiffusionConfig fields: report_path (str/Path, default None — set to enable the PDF report), timestamp_column (str, None), ground_truth_column (str, None), report_title (str, "Kumo-Anomaly Report"), report_max_pages (int, 10).

Inference CLI reference

Fine-tuning

Fine-tune on your own data. The training CSV should contain mostly normal behavior. Validate the pretrained model on your domain before fine-tuning.

uv run python examples/finetune_example.py \
  --csv /path/to/normal_data.csv \
  --val-csv /path/to/val_data.csv \   # optional; otherwise --val-ratio splits --csv
  --pretrained-model final_model.pth \
  --epochs 20 --batch-size 16 --lr 1e-5 \
  --output-dir artifacts/finetune_my_data

Fine-tuning arguments

Running inference with a fine-tuned checkpoint

results = perform_anomaly_analysis_with_diffusion(
    df=df,
    threshold_strategy="scs",
    model_path="artifacts/finetune_my_data/best_finetuned_model.pth",
    model_config_path="artifacts/finetune_my_data/finetune_config.yaml",
    nsample=15,
)

AutoML (HPO: hyperparameter optimization)

This skill supports AutoML for fine-tuning HPO and labeled inference HPO through tao-skill-bank:tao-run-automl with this model's skill_dir.

Read references/automl.md when the user asks for AutoML/HPO setup, tunable parameters, VirtualEnvSDK setup, config flow, window-length constraints, inference trial scripts, or AutoML result handoff details.

Data requirements

target_dim 18 is the config-file default, not a hard constant — when fine-tuning warm-starts from a published checkpoint (--pretrained-model), the checkpoint's embedded config overrides it (the published final_model.pth was trained at target_dim=40). Check the checkpoint's config rather than assuming 18 when reasoning about PCA/padding behavior for warm-started fine-tuning.

timestamp,sensor_1,sensor_2,sensor_3
2024-01-01 00:00:00,0.42,1.10,-0.33
...

Pass only numeric feature columns to the inference API — it raises ValueError on non-numeric columns rather than dropping them. Use df.select_dtypes(include="number") or drop by name before calling. Fine-tuning handles this via --timestamp-col, --label-col, and --drop-cols CLI args.

Output structure

Inference (examples/quick_example.py):

examples/datasets/
└── kumo-anomaly-results.csv # original columns + Anomaly (0/1) + MAE

Fine-tuning (--output-dir artifacts/finetune_my_data):

artifacts/finetune_my_data/
├── best_finetuned_model.pth     # checkpoint with lowest validation loss
├── final_finetuned_model.pth    # checkpoint from last epoch
├── metrics.json                 # scalar for AutoML: {"val_loss": <best>}
├── epoch_metrics.json           # per-epoch log: [{"epoch": N, "train_loss": …, "val_loss": …}]
└── finetune_config.yaml         # config used during training (for reproducibility)

Model configuration (curriculum_medium.yaml)

Hardware

Known pitfalls