---

name: pytorch-vae description: "Routes PyTorch-VAE tasks to config-driven training and model API workflows for the collection of variational autoencoders." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

PyTorch-VAE

Use this skill for the repository that collects many PyTorch variational autoencoder variants, their configs, and their smoke tests.

Two routes matter most:

• Training/config route for experiment runs, CelebA data layout, TensorBoard logs, checkpoints, and trainer settings.
• Model-reference route for constructor arguments, forward/loss behavior, sampling, generation, and registry lookups.

Start here

• Read references/repo-provenance.md when you need to check whether this skill still matches the current checkout.
• Read references/model-overview.md when you need a fast model-family map.
• Read references/troubleshooting.md for cross-cutting install, backend, config, and runtime failures.
• Open sub-skills/training/SKILL.md for config-driven fitting and data/logging questions.
• Open sub-skills/model-reference/SKILL.md for class signatures, latent-shape details, or synthetic smoke checks.

Install

From a PyTorch-VAE checkout, install the runtime stack with a CUDA-capable torch/torchvision pair, then the repo requirements:

python -m pip install 'pip<24.1' 'numpy<2' 'setuptools<70' six
python -m pip install 'torch==1.13.1+cu117' 'torchvision==0.14.1+cu117' --extra-index-url https://download.pytorch.org/whl/cu117
python -m pip install -r requirements.txt

If you are using a different CUDA wheel tag, keep the torch and torchvision versions matched.

Verified baseline

The repo was inspected with a CUDA-capable Python 3.10 environment using:

• torch 1.13.1+cu117
• torchvision 0.14.1+cu117
• pytorch-lightning 1.5.6
• PyYAML 6.0
• torchsummary 1.5.1
• tensorboard 2.21.0
• numpy 1.26.4
• setuptools 69.5.1
• six 1.17.0

That baseline was verified on an NVIDIA A100 host with CUDA 11.7 available.

Route map

Training / config-driven experiments

Use this route when the user wants to choose a config, run or dry-run a training job, validate CelebA data layout, inspect logging output, or troubleshoot trainer arguments.

Helpful bundled entry point:

• sub-skills/training/scripts/train_from_config.py — validates a config from any checkout root and can run a full fit only when --fit is set.

Read sub-skills/training/references/workflows.md before a real training run.

Model reference / smoke testing

Use this route when the user wants to instantiate a model class, compare constructor kwargs, run a tiny forward/loss check, or test sample/generate behavior for a specific architecture.

Helpful bundled entry point:

• sub-skills/model-reference/scripts/model_smoke.py — instantiates a model from a config, runs synthetic forward/loss checks, and can optionally exercise sample/generate paths.

Read sub-skills/model-reference/references/api-reference.md when you need exact signatures or special-case model behavior.

Minimal check

From the generated skill directory, point the bundled helpers at the checkout you want to inspect and run the smallest smoke that matches the task.

• Training dry-run: python ./sub-skills/training/scripts/train_from_config.py --repo-root /path/to/PyTorch-VAE --config /path/to/PyTorch-VAE/configs/vae.yaml
• Model smoke: python ./sub-skills/model-reference/scripts/model_smoke.py --repo-root /path/to/PyTorch-VAE --config /path/to/PyTorch-VAE/configs/vae.yaml

Use --fit only when you actually want the full training loop.

Common guardrails

• This repo is source-first; it does not expose a packaged install entry point.
• Full training is GPU-oriented and expects the Lightning 1.x API used by the repo code.
• Some model families have special inputs or sample behavior. Read the model-reference route before guessing labels, latent sizes, or sample availability.
• If a config or model name is not in the registry, stop and inspect the bundled references instead of guessing.