nv-generate-vae-finetune

NV-Generate-VAE-Finetune

Purpose

• Used for finetuning the NV-Generate-CTMR MAISI VAE/autoencoder from user-supplied CT or MRI NIfTI training volumes.
• Not for clinical interpretation, regulatory use, or approving synthetic data for production training.
• Upstream currently documents VAE training in train_vae_tutorial.ipynb and provides configs/helpers, but not a scripts.train_vae CLI. This skill does not execute the notebook; it stages the required config/datalist glue locally and uses upstream helper APIs.
• Manifest I/O: inputs are datalist and data_base_dir; outputs are autoencoder_checkpoint, discriminator_checkpoint, and result_json.
• The underlying training contract is the upstream config/env JSON (config_maisi_vae_train.json + environment_maisi_vae_train.json, as used in train_vae_tutorial.ipynb). The wrapper stages those JSON files for you and exposes the most-tuned fields as CLI flags; the sections below document the fields, their defaults, and how to monitor/tune a run.

Instructions

• Read skill_manifest.yaml before changing arguments, side effects, or validation gates.
• Run scripts/run_vae_finetune.py from the Medical AI Skills repo root.
• If a host agent exposes run_script, use run_script("scripts/run_vae_finetune.py", args=[...]); otherwise run the Bash/Python command below.
• Use --preflight first when checking a new datalist; remove --preflight only when the user explicitly wants to launch GPU finetuning.
• For a staged preflight input bundle directory, use BUNDLE/preflight_datalist.json as the datalist and BUNDLE/preflight_dataset as --data-base-dir when those files are present.

Examples

Validate and stage a preflight finetune check from an input bundle (the recommended first step — no GPU, no training). This is the single canonical command; replace INPUT_BUNDLE and OUT_DIR with your paths:

export NV_GENERATE_ROOT="${NV_GENERATE_ROOT:-$HOME/.cache/nvidia-skills/upstreams

For real GPU finetuning and other variations, see Usage below.

Available Scripts

Prerequisites

• An explicit NV_GENERATE_ROOT may point to the caller's local checkout and must contain configs/config_maisi_vae_train.json, scripts/transforms.py, and scripts/utils.py. The result records its current commit.
• If NV_GENERATE_ROOT is unset, the wrapper searches .workbench_data/upstreams/NV-Generate-CTMR.
• CUDA_VISIBLE_DEVICES is optional and can be used to select the GPU for real training.
• Runtime requirements: NVIDIA CUDA GPU for real training, Python packages from the upstream requirements.txt, lpips, and downloaded VAE weights unless using --train-from-scratch.
• Side effects: writes staged configs, checkpoints, TensorBoard logs, and run summaries under the caller-provided --output-dir; may write model caches under the upstream checkout, ~/.cache/huggingface/, and ~/.cache/torch/; may contact https://huggingface.co, https://github.com, and https://download.pytorch.org.
• The datalist is a MONAI-style JSON object with non-empty training[] and validation[] or testing[]. Each entry has an image path relative to --data-base-dir and optional class or modality of ct or mri.

When no local checkout is supplied, create the recommended pinned default checkout once:

1. Config and environment JSON (adapt to your data)

The wrapper copies the upstream VAE config/env JSON from $NV_GENERATE_ROOT/configs, rewrites the fields below, and writes the staged copies under OUT_DIR/workflow/configs/. You normally only set your datalist and data root; the listed CLI flags override individual fields when you need to.

Environment JSON (environment_maisi_vae_train.json):

Training fields (config_maisi_vae_train.json):

--modality (ct or mri, default mri) fills the per-entry class for datalist items missing one. Validation/testing entries are required because the training loop runs a validation pass.

For an end-to-end reference including example data download, see the upstream tutorial train_vae_tutorial.ipynb.

2. Usage (one-line training)

Preflight only:

Preflight bundle input:

GPU finetuning:

Replace PATH_TO_DATALIST.json and PATH_TO_DATA_ROOT with the user's actual paths. Do not use the fixture datalist for real training; it is a preflight-only placeholder.

3. Monitor training (TensorBoard)

The runner writes TensorBoard scalars (per-iteration and per-epoch recons_loss, kl_loss, p_loss, adversarial/real/fake losses, and a validation scale_factor) under OUT_DIR/artifacts/tfevent/autoencoder. Launch TensorBoard against the output directory:

The same per-epoch loss history is also captured in OUT_DIR/artifacts/workflow_summary.json and echoed in the JSON the wrapper prints to stdout (loss_history, best-checkpoint paths, exit_code, stderr_tail).

4. Hyperparameter tuning and common pitfalls

• Reconstructions blurry — raise --perceptual-weight (default 0.3); try --recon-loss l2 if edges look washed out.
• Posterior collapse / over-regularized latents — --kl-weight is intentionally tiny (1e-7); increasing it too much degrades reconstruction.
• Adversarial training unstable — lower --adv-weight (default 0.1) or --lr; a warmup schedule already ramps the LR over the first 20 epochs.
• Out-of-memory — reduce --patch-size (e.g. 48,48,48) and --val-sliding-window-patch-size, keep --batch-size 1, and lower --cache-rate.
• datalist must include non-empty validation[] or testing[] — the validation loop is mandatory; add validation[] (or testing[]) entries.
• Single-GPU only — the runner asserts exactly one CUDA GPU; set CUDA_VISIBLE_DEVICES to pick which one.

5. Evaluate the finetuned VAE

Validation reconstruction loss (lowest-val_weighted_loss epoch) is tracked automatically and the best autoencoder is saved as autoencoder_epochN.pt under OUT_DIR/artifacts/models. To evaluate downstream:

• Compare validation recons_loss/p_loss curves across runs in TensorBoard, and
• Plug the finetuned autoencoder into a diffusion finetune/generation run (e.g. nv-generate-mr-brain-finetune via --trained-autoencoder-path) to confirm latents still decode to usable volumes.

This skill gates file accounting and reconstruction bookkeeping only — image quality and downstream utility must be judged by a domain expert.

Limitations

• Requires a current upstream NV-Generate-CTMR checkout with VAE configs and helper APIs. The skill owns the runner glue and does not depend on the notebook.
• Full training can be expensive and is not deterministic across hardware, CUDA, and package versions.
• The wrapper gates file accounting and command provenance, not anatomical realism, reconstruction quality, or downstream model utility.
• Not for clinical deployment, clinical interpretation, autonomous diagnosis, regulatory submission, or production training-data approval.

Troubleshooting