Generative

PaddleViT GANs

Responsibility and boundaries

Use this route when a request names the PaddleViT GAN module, TransGAN, or Styleformer, or asks to generate/evaluate images from one of those transformer GANs. The route owns:

• the two model families' generator/discriminator contracts and shipped YAML variants;
• the CIFAR10, STL10, CelebA, and LSUN-church LMDB data contracts;
• safe generation/evaluation planning, image range/layout checks, FID/PSNR/SSIM interpretation, and training/evaluation resource boundaries;
• dependency diagnosis for Paddle, Pillow, OpenCV, SciPy, LMDB, matplotlib, and tqdm; and
• the download-free bundled contract smoke at scripts/generative_model_smoke.py.

Do not use this route for classification, detection, segmentation, generic Paddle export, or repository-wide environment operations. Do not port PyTorch weights or download checkpoints, Inception weights, or datasets as part of a safe generation/evaluation request. The source contains weight-porting helpers and external model links, but those are deliberately excluded. Full 50k-sample evaluation and multi-epoch/full-dataset training are expensive and require explicit user authorization, local data, local checkpoints, and a verified budget.

This is a self-contained operating guide. It does not import the original checkout at runtime. The source was inspected at PaddleViT commit 5ac7d89d4fd0e3235d055ff15d5b1b1315499d70; source paths in the evidence list are provenance, not runtime dependencies.

Route a request

1. Identify the family and operation before touching a config:

   • choose TransGAN for the pure-transformer generator and two-scale transformer discriminator;
   • choose Styleformer for the style-vector mapping/synthesis generator and StyleGANv2-style convolutional discriminator;
   • choose generate for images only, eval for FID or metric reporting, and train only after the user explicitly accepts data/GPU/compute cost.

2. Choose the dataset and exact image size. The shipped Styleformer YAMLs map CIFAR10→32, STL10→48, CelebA→64, and LSUN-church→128. The TransGAN YAML is CIFAR10→32. Do not silently reuse a 32-pixel config for a 48/64/128-pixel dataset.

3. Inspect the local data layout with the checklist in references/data-formats.md. An absent optional dependency is a preflight failure, not a reason to create a fake dataset or claim evaluation passed.

4. Run the safe, deterministic synthetic smoke before a source build or a costly run:

   python /path/to/this/skill/scripts/generative_model_smoke.py --help
   python /path/to/this/skill/scripts/generative_model_smoke.py --model all --device auto

   This checks small independent Paddle contracts only. It does not import the checkout, load a checkpoint, open a dataset, contact a URL, or prove historical GAN quality. If Paddle is unavailable, the script reports a clearly marked skip.

5. For a source model build, run from exactly one family directory and expose only that directory on PYTHONPATH:

   cd gan/transGAN       # or gan/Styleformer
   export PYTHONPATH="$PWD:${PYTHONPATH:-}"

   Both families use common module names (config.py, utils.py, generate.py). Mixing both source roots can import the wrong family.

6. Record the final config, source commit, device, seed, local checkpoint provenance, data counts, image range/order, metric feature extractor, and sample limits. If any of those are unknown, report the result as partial.

Source command contract

The source uses short single-dash flags. A safe command is a plan/template, not permission to download or run expensive work:

-cfg PATH             family YAML; recursive BASE entries are supported
-dataset NAME         cifar10, stl10, celeba, or lsun where wired
-batch_size N         per-process batch size
-data_path PATH       dataset root or family-specific image/LMDB directory
-image_size N         override only when model architecture supports it
-eval                 evaluation-only path
-pretrained PREFIX    source training/eval appends .pdparams in main scripts
-resume PREFIX        source expects PREFIX.pdparams and PREFIX.pdopt
-num_out_images N     generation count for generate.py
-out_folder DIR       generated image destination

generate.py has family-specific conventions. TransGAN's generator loads a state dictionary under gen_state_dict; Styleformer's standalone generator loader expects the saved generator dictionary directly, while the training scripts expect a combined gen_state_dict/dis_state_dict checkpoint. Verify the actual local file keys before loading. Do not add .pdparams to a -pretrained prefix when using the training/evaluation scripts; generation examples are inconsistent and may accept a complete filename. Prefer an explicit local-file check and a tiny load test over guessing.

Generation and evaluation procedure

Generation

• Set the device before model construction. Use one visible GPU for a normal source generation run; CPU is suitable only for parser/shape diagnostics and may be impractical for the full transformer.
• Seed Paddle (and NumPy/Python if used by the caller) before sampling. The source scripts use random latent tensors and do not promise a stable benchmark seed.
• TransGAN samples z with shape [B, 256], calls Generator(z, epoch), and returns [B, 3, 32, 32] for the shipped CIFAR10 config. Styleformer samples z with shape [B, 512]; its c_dim is zero in the shipped configs, so the mapping network ignores labels even though eval code supplies a label tensor. It returns [B, 3, H, W] with H/W set by the chosen config.
• Both source generation scripts postprocess model output with (output * 127.5 + 128).clip(0, 255).astype('uint8'), then transpose from CHW to HWC and write RGB PNGs. This assumes the model output is in a [-1, 1]-like range. TransGAN's final source Conv2D has no explicit tanh; inspect output statistics rather than treating the assumption as a proven invariant.
• Keep generated tensors in NCHW for Paddle discriminators and metric feature extraction. Convert to HWC only at the image-file boundary.