Chest pneumonia 256 — StyleGAN2-ADA
Unconditional StyleGAN2-ADA trained from random initialization by Purino. The model pools pneumonia and normal training images. It cannot generate a requested class label. Outputs are synthetic research images, not diagnostic evidence.
Source: https://github.com/NVlabs/stylegan2-ada-pytorch/tree/d72cc7d041b42ec8e806021a205ed9349f87c6a4 Dataset: https://www.kaggle.com/datasets/thomasdubail/chest-pneumonia-256x256 Base weights: none (training from scratch).
Training exposure: 1500.032 kimg across sessions; current session: 300.032 kimg. One kimg is 1,000 real-image presentations. This is the latest checkpoint, not a model selected by minimum validation FID.
Two NVIDIA T4 GPUs; FP16 at high resolutions; global batch 64; microbatch 16 per GPU; R1 gamma 1.0; ADA target 0.6. Data are converted to RGB without resizing; exact pixel duplicates are removed. The attached dataset summary records counts. Supplied test/validation folders are excluded; publisher-described patient grouping is not independently verified. FID/KID against this training set are descriptive only.
Recovery uses model.pkl plus training_state.pt from the same Hub commit. It restores G, D, G_ema, Adam states, cumulative counters, ADA probability/statistics, path-length running mean and per-rank random states. Data sampler position and prefetched batches restart; this is not a bit-for-bit replay. Training quality must be assessed from generated samples and downstream evaluation.
Generate samples
Every uploaded checkpoint is kept in its own folder under
checkpoints.
For example, checkpoints/kimg-000010.048/ is the save at 10.048 kimg.
Each folder contains model.pkl, training_state.pt, run_state.json, samples.png
when available, and the configuration/source files from that upload.
The root model.pkl and training_state.pt remain the latest checkpoint for resume.
Choose any checkpoint folder to compare samples manually using the same seeds.
Extract source.zip, install the notebook's pinned environment, and run:
python generate.py --network=/path/to/model.pkl --outdir=generated --seeds=0-15 --trunc=1
model.pkl contains G, D and G_ema. For inference, use G_ema. Load only trusted
pickle files. source.zip includes the runtime compatibility edits and their diff.
Helper scripts are included; the Kaggle notebook is included when exposed as a local file by Kaggle.
Terms and limitations
NVIDIA's complete source license is included in LICENSE.txt; its use limitation is non-commercial research/evaluation. The dataset description specifies CC BY 4.0. No claim of clinical validity, privacy preservation, class fidelity or convergence is made.