Wasserstein-GAN throwaway experiment (fast single-pass eval) #3

New Issue

2026-07-08T13:52:18+02:00

lbogner commented

2026-07-08 13:52:18 +02:00

(Migrated from gitlab.etp.kit.edu)

Try a Wasserstein-GAN variant of the step surrogate as a quick throwaway experiment. The outcome is genuinely uncertain — the point is to measure it.

Why: a GAN generator is a single forward pass, so it's very fast at evaluation — directly relevant to the ~10× native-Geant4 execution-time target.

Context / caveat: GANs were previously ruled out on mode-collapse / hard-zero-tail grounds (generative-architecture-alternatives §7) and a WGAN was deferred at the 2026-06-25 meeting. This is a deliberate empirical re-test, not a reversal — Wasserstein loss is specifically the variant that mitigates mode collapse, so it's the right GAN to probe the tail behaviour (photon edep multimodality, direction concentration) with. Worth comparing both eval speed and tail quality against the Phase 2 baseline.

Source: Jan & Tobias meeting 2026-07-07 §2. Low-effort, runs in parallel with the baseline.

Try a **Wasserstein-GAN** variant of the step surrogate as a quick throwaway experiment. The outcome is genuinely uncertain — the point is to measure it. **Why:** a GAN generator is a *single forward pass*, so it's very fast at evaluation — directly relevant to the ~10× native-Geant4 execution-time target. **Context / caveat:** GANs were previously ruled out on mode-collapse / hard-zero-tail grounds (`generative-architecture-alternatives` §7) and a WGAN was deferred at the 2026-06-25 meeting. This is a deliberate *empirical* re-test, not a reversal — Wasserstein loss is specifically the variant that mitigates mode collapse, so it's the right GAN to probe the tail behaviour (photon `edep` multimodality, direction concentration) with. Worth comparing both eval speed and tail quality against the Phase 2 baseline. Source: Jan & Tobias meeting 2026-07-07 §2. Low-effort, runs in parallel with the baseline.

lbogner commented

2026-07-22 10:35:32 +02:00

(Migrated from gitlab.etp.kit.edu)

created branch 3-wasserstein-gan-throwaway-experiment-fast-single-pass-eval to address this issue

created branch [`3-wasserstein-gan-throwaway-experiment-fast-single-pass-eval`](/lbogner/giant/-/compare/master...3-wasserstein-gan-throwaway-experiment-fast-single-pass-eval) to address this issue

lbogner commented

2026-07-22 10:35:38 +02:00

(Migrated from gitlab.etp.kit.edu)

mentioned in merge request !6

lars referenced a pull request that will close this issue

2026-07-23 15:32:07 +02:00

Draft: Resolve "Wasserstein-GAN throwaway experiment (fast single-pass eval)" #14

lars closed this issue

2026-07-23 15:32:49 +02:00

Sign in to join this conversation.