RMDig/snowGAN-magnified-profile

snowGAN — magnified_profile backbone (v0.1.0)

WGAN-GP trained on the Rocky Mountain Snowpack dataset, single-modality (magnified_profile), depth=1, 1024x1024 resolution. Published from the snowGAN project as a research artifact.

Not for safety decisions. Nothing here predicts avalanche danger, slope stability or snowpack breakability. Do not use it for any backcountry decision.

How to use

The canonical consumer-side path uses snowgan.weights.fetch to download + cache, then snowgan.models.Discriminator to rebuild and load:

from snowgan.weights import fetch
from snowgan.config import build
from snowgan.models.discriminator import Discriminator

# Fetch the release (cached locally after first call).
path = fetch("RMDig/snowGAN-magnified-profile", "v0.1.0")

# Reconstruct the model from its sidecar config, then load weights.
cfg = build(str(path / "discriminator_config.json"))
disc = Discriminator(cfg)
disc.model.build((None, cfg.depth, cfg.resolution[0], cfg.resolution[1], cfg.channels))
disc.model.load_weights(str(path / "discriminator.weights.h5"))

# Tap the named features layer — the cross-repo contract with downstream consumers.
features = disc.model.get_layer("features")
print("backbone features:", features.output.shape)  # (None, 1048576)

Requires pip install snowgan[hub] (pulls in huggingface_hub). Without the [hub] extra, fetch() raises a clean ImportError naming the missing dep.

This release's generator_config.json predates the generator's architecture fields. With snowgan 1.0.0, rebuild the generator with gen_upsampler="transpose" and gen_convs_per_resolution=1; the 1.0.0 defaults describe a different generator and the weights will not load.

Intended use

Research artifact from snowGAN's GAN-training experiments. The discriminator's features layer and the generator are published so the results can be reproduced and tested. Publishing them does not mean they are known to be useful downstream.

Transfer value is unmeasured here and, on the one task tested, worse than random init. In snowGradient's modality probe a randomly initialised copy of this network beat the pretrained backbone (see Quality evidence). Compare against that baseline before relying on these features.

Architecture

Field	Value
Modality	magnified_profile (depth=1)
Resolution	1024x1024
Channels	3
Latent dim	100
Generator filter counts	[1024, 512, 256, 128, 64]
Discriminator filter counts	[64, 128, 256, 512, 1024]
Conv kernel / stride	[3, 3] / [2, 2]
Backbone (Flatten "features") dim	1048576
Final activation	tanh

The discriminator's Conv3D layers use ksize=(1, kH, kW), so the depth axis is broadcast — kernels themselves are depth-agnostic. This is the contract that lets downstream consumers compose multiple single-modality backbones into a depth=N model (e.g. profile + core stacked at depth=2 for paired-modality transfer).

Training

Trained with WGAN-GP loss (Wasserstein + gradient penalty, λ_gp=1.0) on the magnified_profile samples of rmdig/rocky_mountain_snowpack. At release: fade_step=428139.

The settings below are the run's terminal configuration, not the configuration it learned under. Per-snapshot configs, sorted by step, show:

Steps (snapshots)	disc:gen	λ_gp	Spectral norm	Other changes
146.4k-154.4k, 279.6k-288.6k	2:3	10.0	off	none of the stabilizers below
289.6k	2:3	10.0	on	cosine LR, augmentation, EMA on
295.6k	2:3	1.0	on
296.6k	2:3	1.0	on	multiscale critic, ADA, grad clip, adaptive steps on
298.6k → 425.6k	1:3 rising to 61:3 at 425.6k	1.0	on
release (428.1k)	47:3	1.0	on

No config snapshots survive for steps 0-146k or 154k-279k. Before 2026-03-13 (commit 2817926) the gradient penalty was multiplied by λ twice, so a configured λ_gp of 10 applied an effective weight of 100 for the part of the run trained before that date.

Cosine decay was enabled past its then hard-coded 200k horizon, so both main learning rates dropped straight to lr_min = 1e-7. Generator weights moved 12.2% (relative L2) in the 1k steps before 288.6k, and 2.4% in total over the 136k steps from 289.6k to 425.6k. The model acquired its structure at disc:gen 2:3, configured λ_gp 10, without spectral norm; the stabilizers in the table below arrived after learning had largely stopped.

Stabilizer	Setting
Spectral norm	True
Differentiable augmentation	True
Adaptive augmentation (ADA) target	0.6
Adaptive disc/gen step ratio	True
EMA decay (generator shadow)	0.999
Multi-scale discriminator	True
Gradient clip (global norm)	1.0
LR decay schedule	cosine (lr_min=1e-07)
FID eval interval	5000 steps (no valid FID; see Quality evidence)

Training data and splits

• 13 (site, column, core) groups, all from sites 0-2, one operator. The dataset's sites 3-6 were added on 2026-08-17, after this release, and are unseen by it.
• The sidecars persist a trained_pool / validation_pool / test_pool split (10 / 1 / 2 groups). These pools are not held out from this model. The training loop at release ignored them, so the GAN trained on all 13 groups (snowGAN UPGRADES #54).
• The GAN streamed the dataset's row-level train split as it stood at release: 1,101 of the 1,355 images of this modality, covering all 13 groups. The remaining images come from the same groups, so they are not independent held-out data. Sites 3-6 are the only data unseen at the group level.

Quality evidence

• No valid FID. The in-training FID used 64-256 samples against 2048-dimensional Inception features, a rank-deficient and sample-size-biased estimate. The fid_interval setting above produced no usable number.

• KID: the generator is far from real images (measured 2026-10-02/03). Unbiased polynomial KID on Inception features, 200 generated samples, 10 subsets of 100. The pass criterion was fixed before scoring: KID within 2 standard errors of a real-vs-real floor.

  	KID ± SE
  Floor: real vs real, sites 3-6 (two group-disjoint halves, 995 images, 32 groups)	0.0025 ± 0.0005
  Site shift alone: real sites 0-2 vs real sites 3-6	0.069 ± 0.002
  generator.weights.h5 vs sites 3-6 (never seen)	0.386 ± 0.003
  generator.weights.h5 vs sites 0-2 (its training sites)	0.384 ± 0.003
  generator_ema.weights.h5 vs sites 3-6	0.338 ± 0.002
  generator_ema.weights.h5 vs sites 0-2	0.346 ± 0.002

  Both fail: the gap is 62x (primary) and 85x (EMA) the bound. The generator is as far from the images it trained on as from unseen sites, so the gap is the generator, not the change of site.

• No memorisation. For 200 samples per weights file, the mean distance to the nearest training image (128 px, mean absolute difference) is 1.03x (primary) and 1.08x (EMA) the distance from a training image to the nearest image of a different group. Samples are not copies. Only 86 / 95 of 200 samples have distinct nearest neighbours, consistent with limited variety.

• The logged critic loss never exceeded the 1-Lipschitz ceiling. Max |logged loss| was 783 over 427,220 steps, against 2sqrt(10241024*3) ≈ 3,547. That is necessary for a 1-Lipschitz critic, not proof of one.

• The generator is not collapsed. snowGAN's own check (2026-09-23, raw weights, 6 samples) measured latent diversity 0.490, against ≤ 0.00003 for collapsed runs. It produces varied images that do not copy its training set, but by KID they are not close to real snow images.

• Downstream measurement (snowGradient, 2026-10). In a pre-registered modality-recognition probe (core vs profile vs magnified profile, leave-one-site-out over sites 3-6, at 256x256 with global average pooling), the pretrained magnified-profile backbone scored 0.868 balanced accuracy. A randomly initialised copy of the same network scored 0.987 / 0.982 (2 seeds) and was better on every held-out site. 24 of the pretrained backbone's 26 errors were core <-> magnified-profile confusions. This is one task, at a quarter of training resolution.

• Its features appear specialised to magnified-profile texture, given that error pattern.

Limitations

• Trained exclusively on the magnified-profile modality of the rmdig/rocky_mountain_snowpack dataset; generalization to other distributions untested.
• Single-modality backbone (depth=1); paired-modality features must be composed on the consumer side. See AvAIs load_backbones for the canonical two-backbone composition pattern.
• 13 groups from sites 0-2, one operator: a very small, narrow training set. There is no held-out data from those sites (see Training data and splits).

Release history

Two commits titled "Release v0.1.0" were made about 50 minutes apart on 2026-06-03. The first carried no model card (a 3-line README); the second added this card and corrected the MANIFEST's provenance, and filled in the discriminator sidecar's split pools. The v0.1.0 tag was created on the second commit, one second after it; there is no record of a tag on the first. All weight files are byte-identical across both commits (same LFS sha256), so either snapshot computes the same function.

Card corrected 2026-10: removed unmeasured claims about feature quality and transfer learning, and added the measurements above.

Files in this release

• discriminator.weights.h5 — main discriminator weights.
• discriminator_config.json — architecture sidecar; pass to snowgan.models.Discriminator(cfg).
• generator.weights.h5 + generator_config.json — generator weights and sidecar.
• generator_ema.weights.h5 — EMA shadow weights (only if EMA was enabled during training).
• generator_fade_endpoints.weights.h5 — progressive-fade toRGB endpoints (only if fade was used).
• discriminator_lowres.weights.h5 — multi-scale 256×256 critic (only if multiscale_disc was on).
• MANIFEST.md — full provenance dump (git SHA, every training flag, every artifact). Read this for debugging.
• README.md — this file.

License

Apache 2.0 — see the snowGAN repository for the full license text.

Cross-references

• Source code: github.com/dennys246/snowGAN
• Downstream project: github.com/dennys246/AvAI
• Companion modality backbones: RMDig/snowGAN-magnified-profile, RMDig/snowGAN-core
• Training dataset: rmdig/rocky_mountain_snowpack

Release notes

First public release. Trained ~428k steps (427,220 logged). Spectral norm, augmentation, EMA, gradient clipping, cosine LR and adaptive steps were switched on only after ~289k steps, by which point learning had largely stopped (see Training). The logged discriminator loss stayed far below the Lipschitz ceiling throughout.

Corrected 2026-10. Earlier text said the model was trained "with full production stabilizers", which describes only the last ~130k steps. It also said this release "supersedes an earlier v0.1.0". See Release history: the earlier commit had identical weights and no card.