Figure 6 — Pool generator (β-VAE)

Vector PDF

File: fig06_pool_generator_vae.pdf

Suggested caption. The window-pool generator. A convolutional β-VAE is trained on core training windows and stopped early on the holdout loss. After training, N windows are sampled from a shrunk prior and written to disk as the pool that every same-pool method shares.

Details

Item	Value
Architecture	tsgm vae_conv5 encoder/decoder + BetaVAE, latent dimension 16
Loss	reconstruction + β·KL, β = 2
Optimisation	Adam, lr 1e-3, batch 64, up to 400 epochs
Early stopping	holdout loss after each epoch with z = μ; patience 50; best weights restored; epochs 1–150 are not eligible (vae_min_epochs; taken from earlier work of the authors on other data)
Sampling	z ~ 0.5 · N(0, I) → decoder → NaN replaced by 0.5 → clipped to [0, 1]
Pool	N = max(300, 2K) windows of shape (T, 3), per (dataset, seed)

Model sharing across seeds (pool_train_replicas)

• min: one VAE per dataset (9 trainings in total); seeds change only the sampling.
• mid, max: one VAE per seed (90).
• Any other value k of pool_train_replicas gives k VAEs per dataset; seed s uses model s mod k.

The training time of a shared model is recorded for every seed and enters the cost comparison (RQ6) for each of them.

Notes for the text

• The assembler does not depend on where the pool comes from. Any generator that outputs windows in the same three channels can replace the VAE (registry pool_generators; e.g. sdv_par, real_windows).
• Fitting and sampling have separate fingerprints. Changing N re-samples the pools from the saved weights without retraining the VAE.