lars/giant
1
0
Fork 0
Code Issues 4 Pull Requests 2 Actions Packages Projects Releases Wiki Activity
86 Commits 5 Branches 1 Tag
36bdbf7cc174df12977a458e3a07580ec05bbc21
Commit Graph

5 Commits

Author SHA1 Message Date
lars e6e0eb22bf Implement Phase 2: secondary particle prediction 
Two-stage factorisation: Stage 1 predicts 9D primary kinematics + n_sec
classification head (COND_DIM reduced to 8, dropping n_sec/e_sec inputs);
Stage 2 (SecondaryDecoder) generates K_MAX=15 secondary slots via masked
flow matching over (stick_logit, local_dir, type_emb) conditioned on Stage 1
output. Joint training with combined loss L_s1 + λ_nsec*L_nsec + λ_s2*L_s2.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-06-29 11:34:31 +02:00
lars 8475199609 Encode edep/secondary/post energy as a conservation-constrained simplex 
Replaces the independent log_delta_e/log_edep targets with 2 additive-log-ratio
coordinates over the deposit/secondary/post-energy simplex (fractions of pre_E
summing to 1), so edep + e_sec + post_E == pre_E holds by construction after
decoding (softmax) rather than being learned approximately. Requires e_sec
(secondary energy) as a new conditioning input and a steps_to_parquet.py pass
to derive it from child track first-step energies.
 2026-06-25 16:01:13 +02:00
lars 53fd2e4405 Add ruff and ty as dev dependencies, fix lint/type findings 
ruff removed unused imports across analysis.py and several test files.
ty caught a wrong dict[int, int] annotation on StreamingStepsDataset's
mat_map (materials are strings) and a real bug in steps_to_parquet.py
where --compression none passed None to polars' write_parquet, which
only accepts the literal "uncompressed". Also narrows a few
Optional-typed attributes (ddpm_schedule, Normalizer.mean/std) with
asserts and aligns __getitem__'s parameter name with torch's Dataset
base class.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-06-18 17:38:40 +02:00
lars 72bd65ff9f Add post_pos as a model target via travel_dir decomposition 
step_length already encodes |post_pos - pre_pos| by definition, so a raw
post_pos target would duplicate that magnitude and could drift inconsistent
with step_length during sampling. Instead add travel_dir, a unit vector
(local frame) giving only the direction of pre_pos->post_pos; post_pos is
reconstructed at inference as pre_pos + step_length * travel_dir, keeping
the two self-consistent. Target grows from 6D to 9D.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-06-18 10:36:55 +02:00
lars 9277d79dff Implement Phase 1: full data pipeline, model, training, and config support 
- Data pipeline: loader (parquet→numpy), transforms (log, local-frame
  Rodrigues rotation, Normalizer), StepsDataset with event-ID-based split
- Model: SinusoidalEmbedding, ConditionEncoder, ResBlock, DenoisingMLP
- Schedule: cosine DDPM and conditional flow matching loss (Lipman 2022)
- Samplers: flow (Euler ODE), DDPM ancestral, DDIM deterministic
- Training loop: AdamW + cosine LR, grad clipping, best-val checkpoint
- Validation: per-dimension marginal summary (normalised space)
- CLI: TOML config support with CLI-overrides; hyperparam-encoded output
  directory; config.toml with git hash saved into each run's checkpoint dir
- 21 unit tests covering transforms, network, flow/DDPM losses, dataset splits

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-06-17 10:48:03 +02:00