| # snowGAN release v0.1.0 |
|
|
| Snapshot of a snowGAN training run, packaged for downstream consumers |
| (AvAI etc.). Consume with `snowgan.weights.fetch("RMDig/snowGAN-core", "v0.1.0")`. |
|
|
| ## Provenance |
| - snowGAN package version: `unknown` |
| - snowGAN git SHA at release: `fde5671fed1b746962b6ca381e3a9b20ccf62e1c` |
| - Training dataset: `rmdig/rocky_mountain_snowpack` |
| - Source save_dir: `/mnt/d/GitSpot/snowGAN/keras/snowgan/core` |
| |
| ## Model architecture |
| - depth: 1 |
| - resolution: [1024, 1024] |
| - modality: core |
| - channels: 3 |
| - latent_dim: 100 |
| - filter_counts (gen): [1024, 512, 256, 128, 64] |
| - filter_counts (disc): [64, 128, 256, 512, 1024] |
| - kernel_size / kernel_stride: [3, 3] / [2, 2] |
|
|
| ## Training state at release |
| - fade_step: 130000 (gen) / 130000 (disc) |
| - fade_steps target: 50000 |
| - current_epoch: 0 |
| - training_steps: gen=3, disc=2 |
| - lambda_gp: 10.0 |
| |
| ## Advanced training options |
| - spectral_norm: False |
| - augment: True |
| - multiscale_disc: True |
| - ema_decay: 0.999 |
| - lr_decay: cosine (lr_min=1e-07) |
| - ada_target: 0.6 |
| - adaptive_steps: False |
| - grad_clip_norm: 1.0 |
| - fid_interval: 5000 |
| |
| ## Persisted dataset splits |
| - trained_pool: 10 groups |
| - validation_pool: 1 groups |
| - test_pool: 2 groups |
|
|
| ## Artifacts |
| - `MANIFEST.md` |
| - `discriminator.weights.h5` |
| - `discriminator_config.json` |
| - `discriminator_lowres.weights.h5` |
| - `generator.weights.h5` |
| - `generator_config.json` |
| - `generator_ema.weights.h5` |
| - `generator_fade_endpoints.weights.h5` |
|
|
| ## Notes |
| First core release. Trained ~127k steps without spectral_norm (the v0.1.0 ms2 fade-step counter shows higher numbers but real training is ~127k). Disc loss diverged post-step-80k due to small dataset (~13 unique cores) + no Lipschitz constraint + unconstrained multiscale_disc lowres critic. Generator produces structured outputs (vertical snow-like patterns, blue/white palette) so the Conv3D backbone has learned meaningful features even if the final Dense head is noisy. Backbone usable for transfer learning; v0.2 planned with spectral_norm enabled for a stable retrain. |
| |
