---

name: biggan-pytorch description: "Guides CUDA-based BigGAN-PyTorch training, checkpoint sampling, dataset preparation, metric preparation, model customization, and optional TensorFlow Hub weight conversion." disable-model-invocation: true metadata: disco-role: operating license: MIT

BigGAN-PyTorch

Use this skill for the author's unofficial PyTorch implementation of BigGAN and its adjacent BigGAN-deep, CIFAR, ImageNet/HDF5, sampling, metric, and TFHub conversion workflows. This is a research codebase rather than an installable library: run its entry points from a checkout, keep outputs outside source files, and treat long training and downloads as explicit user-approved work.

Route the request

• Training, fine-tuning, architecture or optimizer changes, checkpoint resume, multi-GPU, EMA, spectral norm, or mixed precision: read sub-skills/training/SKILL.md and its references.
• Generate images, load a checkpoint, use EMA, standing statistics, truncation curves, NPZ export, sample sheets, IS, or FID: read sub-skills/sampling/SKILL.md.
• Prepare ImageFolder/CIFAR data, convert ImageNet to HDF5, validate HDF5, or calculate Inception moments: read sub-skills/data-preparation/SKILL.md.
• Port DeepMind TFHub BigGAN weights: read sub-skills/tfhub-conversion/SKILL.md; this route is reference-only unless legacy TensorFlow 1.x, TensorFlow Hub, network access, and compatible GPU support are deliberately provisioned.

Read references/model-overview.md for the verified model/data matrix and references/troubleshooting.md for cross-cutting failures. Read references/repo-provenance.md before relying on version-sensitive details or planning a refresh.

Environment and invocation

The repository has no packaging metadata or console entry point. Use a Python environment with a CUDA-capable PyTorch/torchvision pair plus NumPy, SciPy, h5py, Pillow, and tqdm. The README documents an old PyTorch 1.0.1 baseline; the core modules were import-checked and a tiny CUDA generator forward pass was verified with a modern compatible PyTorch installation. Do not assume every legacy option works unchanged on a new PyTorch release.

Set SKILL_ROOT to the directory containing this file and REPO_ROOT to the checked-out BigGAN-PyTorch repository. Repository entry points run from REPO_ROOT; bundled skill helpers are invoked through their SKILL_ROOT paths:

SKILL_ROOT=/path/to/skills/disco/biggan-pytorch
REPO_ROOT=/path/to/BigGAN-PyTorch
cd "$REPO_ROOT"
python "$SKILL_ROOT/scripts/check_environment.py" --repo-root "$REPO_ROOT"

Other repository entry points can then be run from the same checkout, for example:

python -c "import torch; print(torch.__version__, torch.cuda.is_available())"
python train.py --dataset C10 --num_epochs 1 --batch_size 8
python sample.py --dataset C10 --experiment_name <name> --load_weights best0

The entry points hard-code cuda in training, sampling, metric, interpolation, and several utility paths. A CPU import is useful for static inspection but is not evidence that the operational workflows work. The environment helper above is intentionally diagnostic; it does not download data, modify checkpoints, or launch training.

Shared operational rules

1. Decide the dataset and resolution first. Dataset names (I32, I64, I128, I256, their _hdf5 variants, C10, and C100) determine image size, class count, root name, and model output shape.
2. Keep data_root, weights_root, logs_root, and samples_root explicit. base_root can re-root the four locations, but parent directories must already exist when the code creates child directories.
3. Use the launch recipes as parameter references, not as unattended jobs. ImageNet preparation, TFHub downloads, full sampling metrics, and training are expensive or network-dependent.
4. Record the exact model module (BigGAN or BigGANdeep), latent settings, class count, EMA choice, and checkpoint suffix with every result.
5. Do not compare this repository's PyTorch Inception scores/FID directly with official TensorFlow metrics; read the sampling and metric references.

The bundled scripts are safe diagnostics or validators. They do not replace the repository's full training/data entry points; pass an explicit --repo-root when a diagnostic needs to import the checked-out model. They never download data, mutate checkpoints, or launch a production run by default.