Add the cosine OGSE class cos-f50-T20 (50 Hz, one 20 ms period each side of the 180, edges 200 T/m/s as built at b = 1000) from the quantised replay pack; group g0's tiers (#2)

- Add the cosine OGSE class cos-f50-T20 (50 Hz, one 20 ms period each side of the 180, edges 200 T/m/s as built at b = 1000) from the quantised replay pack; group g0's tiers (7a01c4b8c618cb79b511f0cf6daf7c944d2e58f6)

README.md

@@ -10,21 +10,23 @@ tags:
 
 # DiSCo time segments
 
-The DiSCo replay walk, contracted at DiSCo's own seven gradient timings so that a whole-grid image costs one pass
-over 125 bytes per walker. It comes from the same 321,138,027 walkers as the replay pack in
-[SubstrateCommons/disco-replay](https://huggingface.co/datasets/SubstrateCommons/disco-replay). It was contracted
-from the walk's unquantised coefficients beside each shard, not from the pack.
+The DiSCo replay walk, contracted at eight gradient waveforms -- DiSCo's own timings, the replay paper's clinical,
+research and Connectome 2.0 timings, a stimulated echo and a cosine OGSE -- so that a whole-grid image costs one pass
+over 133 bytes per walker. It comes from the same 321,138,027 walkers as the replay pack in
+[SubstrateCommons/disco-replay](https://huggingface.co/datasets/SubstrateCommons/disco-replay). The seven square-lobe
+classes were contracted from the walk's unquantised coefficients beside each shard; the OGSE class from the
+quantised replay pack, its quantisation charged in its own certificate term.
 
-**Choose the dataset by the question.** Use this one for an acquisition at one of the timings below, where every
-other knob stays free. Use the replay pack for any other waveform: another δ or Δ, OGSE, an arbitrary gradient.
+**Choose the dataset by the question.** Use this one for an acquisition on one of the waveforms below, where every
+other knob stays free. Use the replay pack for any other waveform: another δ or Δ, another OGSE, an arbitrary gradient.
 
 ## What a timing class leaves free
 
-Each class fixes the times at which the gradient may switch; its segments are constant-gradient intervals. Within
-a class, everything else is chosen at replay time:
+Each class fixes the waveform's shape in time: a square-lobe class's terms are constant-gradient intervals, the OGSE
+class's two terms are its cosine blocks as stored. Within a class, everything else is chosen at replay time:
 
 - the b-value and direction of every measurement, up to the class's gradient amplitude;
-- the per-voxel delivered gradient (gradient nonlinearity, a background gradient);
+- the per-voxel delivered gradient (gradient nonlinearity, a background gradient; square-lobe classes);
 - T₂ and T₁ per pool, and surface relaxivity ρ₂ ≤ 3 µm/s;
 - B₀ of 1.5, 3 or 7 T in any direction, and myelin χ up to 2 × 10⁻⁷, isotropic and anisotropic.
 
@@ -37,10 +39,11 @@ a class, everything else is chosen at replay time:
 | `d17-D30` | δ 17, Δ 30 | spin echo, TE 53.5 |
 | `d17.7-D35.8` | δ 17.7, Δ 35.8 | spin echo, TE 53.5 |
 | `ste-d7.6-TM38.3` | δ 7.6, mixing time 38.3 | stimulated echo |
+| `cos-f50-T20` | cosine 50 Hz, one 20 ms period each side of the 180, edges ramped at 200 T/m/s as built at b = 1000 s/mm² | spin echo, TE 53.5 |
 
 Each class's widths are derived at the 1.003 % floor. The layout's own error bound is in `disco/manifest.json` under
-`certificate`, term by term. Against the replay pack, 16 voxels in all seven classes agree to 2.2 × 10⁻⁵, against
-a combined bound of 3.0 × 10⁻³.
+`certificate`, term by term. Against the replay pack, 16 voxels in the seven square-lobe classes agree to 2.2 × 10⁻⁵,
+and in the OGSE class to 2.2 × 10⁻⁴ bare and 4.5 × 10⁻⁴ with every tier at 3 T, against a combined bound of 3.0 × 10⁻³.
 
 ## Use me
 
@@ -63,10 +66,10 @@ wm = Tissue(T2={"intra": 0.05, "extra": 0.055, "myelin": 0.01}, rho2=1e-6, chi_i
 S_3T, _ = layout.image(seq, tissue=wm, scanner=3.0)          # relaxation, wall contact and the sheath field at 3 T
 ```
 
-`layout.fit(seq)` names the class a sequence fits, or says why it fits none. `partitions=None` fetches every
-class (40 GB). A whole-grid image keeps the class's
-columns in memory, about 9 GB for bare diffusion and more with the tiers; `resident=False` frees them after the
-call.
+`layout.fit(seq)` names the class a sequence fits, or says why it fits none; an OGSE row is the class's stored shape
+scaled to its b (`layout.partitions["cos-f50-T20"].shapes`). `partitions=None` fetches every class (43 GB).
+`resident=False` streams the packed words to the device a chunk at a time: one class of the whole grid with every
+tier in about 3.4 GB (dmipy-sim#823); `resident=True` keeps them there across calls (about 21 GB with the tiers).
 
 ## Files
 
@@ -74,7 +77,7 @@ call.
 |---|---|
 | `disco/manifest.json` | grid, classes and their segments, widths, certificate, provenance |
 | `disco/w.npy`, `disco/tiles.npy`, `disco/pool.npy` | each row's weight, its voxel half, and its pool |
-| `disco/p_<class>.npy` | each class's segment vectors, packed |
+| `disco/p_<class>.npy` | each class's term vectors, packed |
 | `disco/t_g0.npy`, `disco/t_g1.npy` | the relaxation, contact and field tiers of the spin-echo and stimulated-echo groups |
 | `disco/bg_g0.npy`, `disco/bg_g1.npy` | the background-gradient term of each group |
 

disco/p_cos-f50-T20.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b494163e97ceab755a4b0feb0ade9a2381f267a08f56f5ca5c4bef2edb5a55df
+size 2587748480