Skip to content
Back to skills

Vae

ASecurity

"Routes legacy VAE-family MNIST scripts, backend choices, model

  • 247 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 9, 2026
datagobackend

Works with

  • cli

Security analysis

A100/100

Pro scans all 3 files and shows the line behind each finding

Scanned September 9, 2026

npx -y skills add VectorSpaceLab/AREX-Skill --skill vae --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Vae?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Vae
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/vectorspacelab-vae/badge)](https://www.skillsdirectory.com/skills/vectorspacelab-vae)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: vae
description: "Routes legacy VAE-family MNIST scripts, backend choices, model
  variant lookup, and TensorFlow/PyTorch compatibility troubleshooting."
disable-model-invocation: true
metadata:
  disco-role: operating
license: Unlicense
---

# VAE

Use this sub-skill for VAE-family requests in the Generative Models repository: vanilla VAE, conditional VAE, denoising VAE, adversarial autoencoder, and adversarial variational Bayes.

## Route here when

- The user asks which VAE variant or source artifact label matches a modeling goal.
- The user needs TensorFlow-vs-PyTorch guidance for a legacy VAE example.
- The user asks about MNIST inputs, generated image outputs, latent-variable assumptions, or why a VAE script fails on a modern stack.
- The user names CVAE, DVAE, AAE, or AVB.

## Route elsewhere when

- GAN variants: `../gan/SKILL.md`
- Binary RBM CD/PCD: `../rbm/SKILL.md`
- Helmholtz Machine wake-sleep: `../helmholtz-machine/SKILL.md`
- Shared catalog, provenance, and compatibility: `../../SKILL.md`

## Fast decisions

- Default VAE baseline: vanilla VAE.
- Class-conditioned generation: conditional VAE.
- Denoising / corrupted input reconstruction: denoising VAE.
- Adversarial latent regularization: adversarial autoencoder.
- Adversarial variational objective with a T network: adversarial variational Bayes.

## Shared conventions

- These are standalone legacy MNIST training loops, not a packaged library or CLI.
- The TensorFlow branches use old placeholder/session style.
- The PyTorch branches still rely on the legacy TensorFlow MNIST loader and old scalar logging patterns.
- Sample images are written to a working-directory-local `out/` directory in source checkouts.

## Read next

- `references/workflows.md` for the full VAE variant map and framework coverage.
- `references/troubleshooting.md` for VAE-specific modern-stack failures.
- `../../references/model-catalog.md` for the repo-wide source artifact catalog.
- `../../references/compatibility.md` for the shared TensorFlow/PyTorch/NumPy compatibility matrix.
- `../../scripts/check_legacy_stack.py` before claiming an unchanged VAE example can execute.

Files in this skill

  • SKILL.md2.1 KB
  • references/troubleshooting.md2.2 KB
  • references/workflows.md3.1 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.