NV Brain MRI Diffusion Fine-Tuner
Fine-tune the NV-Generate-CTMR diffusion UNet from NIfTI brain MRI training volumes.
The skill clearly limits use to engineering research, discloses outputs, network endpoints, caches, GPU training, and writes, and records the upstream commit. However, it permits Bash execution, dependency installation, GitHub/Hugging Face access, modification of the user Python environment, and home-directory writes without per-run confirmation, sensitive-MRI handling guidance, isolation, or rollback, so points are deducted.
The entrypoint, configuration layering, input validation, staged command plan, exit codes, and stderr-tail reporting are reasonably clear, and unit-test source is included. However, no independently verifiable execution or CI coverage is supplied; the real path depends on the upstream checkout, CUDA, weights, and package versions, while abnormal-case recovery remains incomplete. The score stays within the static-review ceiling.
The audience, NIfTI/MONAI datalist format, modality choices, preflight mode, and non-clinical boundary are fairly clear. However, the skill targets a CUDA-capable upstream environment, provides no Chinese-language interaction guidance, and depends centrally on GitHub and Hugging Face without explaining mainland-China reachability or alternatives, so points are deducted.
SKILL.md covers purpose, prerequisites, commands, configuration, monitoring, tuning, limitations, and troubleshooting. The manifest supplies versioning, I/O, dependencies, a pinned upstream commit, and side effects; Apache-2.0 licensing and authorship are stated. Missing pieces include a formal changelog, durable maintenance owner/update path, inconsistent version metadata in the skill card, and a reported author-format schema finding.
The wrapper directly supports preflight, config staging, embedding generation, training, optional inference, and result summaries, with a useful preflight-first workflow. However, real training and generation require an external checkout, model weights, CUDA, and user data; the two-epoch default is only for verification, and no statically verifiable representative artifact or quality assessment is provided. The score therefore remains conservative.
The source code, manifest, tests, benchmark report, and pinned upstream revision provide useful audit trails, and outputs record commands, commits, and stderr. However, this review is static, the benchmark is small with no negative tasks, and no independently reproducible CI evidence or real training result is supplied, limiting evidence coverage.
- Before execution, explicitly confirm and isolate Bash execution, dependency installation, network access, home-directory caching, and the caller-provided output path.
- Do not use generated MRI volumes or checkpoints for clinical, regulatory, production-data approval, or unsupervised expert conclusions.
- Verify the upstream checkout, model weights, dependency versions, and network reachability, and establish local storage and log-redaction controls for sensitive imaging data.
What does this skill do, and when should you use it?
This skill fine-tunes the NV-Generate-CTMR rflow-mr-brain diffusion UNet using user-supplied NIfTI training volumes. Its Python wrapper stages configuration, creates latent embeddings, and delegates training to the upstream scripts, with optional inference. A GPU-free preflight check is recommended before launching training. It is not intended for clinical interpretation, regulatory use, autonomous diagnosis, or approval of synthetic data for production training.
Reads a MONAI-style datalist JSON and data root, validating relative image paths and modality metadata; runs scripts/run_mr_brain_finetune.py; stages the configuration and datalist, calls scripts.diff_model_create_training_data to create embeddings, writes embedding sidecars, and calls scripts.diff_model_train; optionally calls scripts.diff_model_infer. It writes embeddings, checkpoints, logs, a workflow summary, and optional inference images under the output directory, and prints result JSON to stdout. It may contact Hugging Face and GitHub for model assets or the upstream checkout.
- Medical AI researchers validating fine-tuning on their own NIfTI brain MRI volumes.
- Model developers adapting the diffusion UNet to supported brain MRI modalities such as mri_t1, mri_t2, or mri_flair.
- Engineers checking datalist paths, modality metadata, and configuration before committing GPU resources to training.
- Research teams producing a fine-tuned checkpoint and reviewing it through built-in inference or the nv-generate-mr-brain inference skill.
What are this skill's strengths and limitations?
- Provides a wrapped workflow from datalist validation through embedding creation and checkpoint output.
- Supports preflight checks, mixed precision, multi-GPU execution, and optional inference.
- Allows tuning learning rate, batch size, cache rate, and inference parameters through configuration JSON.
- Records checkpoints, embedding sidecars, inference outputs, and command results for run tracking.
- Real training requires an NVIDIA CUDA GPU, upstream dependencies, and downloaded MR-brain weights.
- The first run precomputes latent embeddings, which can increase startup time and resource use.
- Training is not deterministic across hardware, CUDA, and package versions.
- The skill manages file accounting and command provenance, not anatomical realism or downstream utility, and provides no clinical or production approval basis.
How do you install this skill?
Install the NVIDIA skill from the collection with: npx skills add nvidia/skills --skill nv-generate-mr-brain-finetune --yes. The README does not document a separate manual installation procedure for this skill; runtime use also requires executing from the Medical AI Skills repository root and preparing an NV-Generate-CTMR checkout.
How do you use this skill?
Run the preflight first: export NV_GENERATE_ROOT="${NV_GENERATE_ROOT:-$HOME/.cache/nvidia-skills/upstreams/NV-Generate-CTMR-61c4ec7}" && python skills/nv-generate-mr-brain-finetune/scripts/run_mr_brain_finetune.py PATH_TO_DATALIST.json --data-base-dir PATH_TO_DATA_ROOT --output-dir runs/nv_generate_mr_brain_finetune_preflight --preflight. After validation, install the upstream requirements, remove --preflight, and add options such as --epochs, --modality, --num-gpus, or --run-inference as needed. Do not use the fixture datalist for real training.