#7 - wai-rs: jpegai_pipeline::reconstruct_yhat — the SGMM autoregressive spine - Transaction-Science/open-standards

dcharlot commented

2026-07-12 22:36:51 -04:00

Owner

Rebuilds the synthesis-input latent y_hat from the dequantized residual + hyper params — the last large structural piece of the decode chain. Ports the reference context.pred loop: chunk psi into 4 per-stage params, down_shuffle the residual into a 2x2 checkerboard (order (0,0),(1,1),(0,1),(1,0) — diagonal pair first), run each MCM stage predicting a mean from the hyper params + already-reconstructed stages (context ASCENDING, y_0 first — pinned from the reference source, not the ONLX input label y[1-0]), add residual, up_shuffle back.

Reuses the byte-exact MCM stages. Validated on real test256.bin data: 76.83 dB SNR (max|Δ|=0.005) vs the reference yhat_y, carrying the MCM nets 16-bit PTQ floor (the float reference loop matches to ~139 dB). Unit test pins the shuffle round-trip + checkerboard order.

Decode chain now structurally complete: entropy → dequant → sigma → reconstruct → synthesis, each validated on real data.

Rebuilds the synthesis-input latent y_hat from the dequantized residual + hyper params — the last large structural piece of the decode chain. Ports the reference context.pred loop: chunk psi into 4 per-stage params, down_shuffle the residual into a 2x2 checkerboard (order (0,0),(1,1),(0,1),(1,0) — diagonal pair first), run each MCM stage predicting a mean from the hyper params + already-reconstructed stages (context ASCENDING, y_0 first — pinned from the reference source, not the ONLX input label y[1-0]), add residual, up_shuffle back. Reuses the byte-exact MCM stages. Validated on real test256.bin data: 76.83 dB SNR (max|Δ|=0.005) vs the reference yhat_y, carrying the MCM nets 16-bit PTQ floor (the float reference loop matches to ~139 dB). Unit test pins the shuffle round-trip + checkerboard order. Decode chain now structurally complete: entropy → dequant → sigma → reconstruct → synthesis, each validated on real data.

dcharlot added 1 commit

2026-07-12 22:36:51 -04:00

wai-rs: jpegai_pipeline::reconstruct_yhat — the SGMM autoregressive spine (real data 76.8 dB) aeffed5f13

Rebuilds the synthesis-input latent y_hat from the dequantized residual + the hyper
params — the last large structural piece of the JPEG-AI decode chain. Ports the
reference context.pred loop:

  p  = chunk(psi, 4)                            # per-stage hyper params, ascending
  r  = down_shuffle(resi) -> (k00,k11,k01,k10)  # 2x2 checkerboard sub-lattices
  for i in 0..4:
      mu_i = MCM_i(p_i, cat(y_0 .. y_{i-1}))    # ctx grows 0/c/2c/3c channels
      y_i  = r_i + mu_i
  y_hat = up_shuffle(y_0 .. y_3)

Reuses the byte-exact MCM stages (mcm_stage0 / mcm_stage_n). Two subtleties pinned
from the reference source (not the ONNX): the checkerboard order is (0,0),(1,1),
(0,1),(1,0) — the diagonal pair first — and the context concatenation is ASCENDING
(y_0 first), despite the ONNX input being labelled y[1-0]; stage 0 has no context.

Validated on real test256.bin data: reconstruct_yhat(resi_dq_y, psi_y) vs the
reference yhat_y = 76.83 dB SNR (max|Δ| = 0.005), carrying the MCM nets' 16-bit PTQ
floor (the float reference loop matches yhat_y to ~139 dB). Unit test pins the
shuffle round-trip + checkerboard ordering.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

dcharlot merged commit 96b431b7fe into main

2026-07-12 22:42:14 -04:00

dcharlot referenced this pull request from a commit

2026-07-12 22:42:15 -04:00

Merge pull request 'wai-rs: jpegai_pipeline::reconstruct_yhat — the SGMM autoregressive spine' (#7) from feat/jpegai-reconstruct into main

dcharlot referenced this pull request

2026-10-21 07:15:17 -04:00

jcp: WithinWindow, a sliding-window caveat over calls, payloads and measured joules #388

dcharlot referenced this pull request

2026-10-21 10:50:11 -04:00

sandbox: flow across a run's checkpoints — an untrusted read closes the sinks that follow #395

dcharlot referenced this pull request from a commit