Docs: note training data/schemes/accels/ACS and how to override at inference

README.md

@@ -15,33 +15,63 @@ Inference-ready `vsharp_<anatomy>.yaml` + `vsharp_<anatomy>.pt` pairs for
 [vSHARP](https://arxiv.org/abs/2309.09954) (Yiasemis et al.).
 
 Configs are **inference-only** (no `training` / `validation` blocks).
+Weights are converted to the current DIRECT ModConv parameter layout.
+
+## What these models were trained on
 
-## Training protocol (important)
+Each anatomy was trained on its own multi-coil dataset with **mixed sampling
+schemes** and **mixed acceleration / ACS** (not a single fixed 4× setting):
 
-These models were **not** trained at a single fixed 4× acceleration or a single
-sampling scheme. Training used:
+| Name | Training data | Default inference mask | Checkpoint iter |
+|------|---------------|------------------------|-----------------|
+| `vsharp_brain` | fastMRI brain (multi-coil) | FastMRIRandom | 950000 |
+| `vsharp_knee` | fastMRI knee (multi-coil) | FastMRIEquispaced | 1050000 |
+| `vsharp_prostate` | fastMRI prostate (multi-coil) | FastMRIEquispaced | 500000 |
+| `vsharp_cardiac` | CMRxRecon 2023 (multi-coil cine) | FastMRIEquispaced | 100000 |
+
+Shared masking schedule used during training (all schemes below):
 
 | | |
 |---|---|
-| **Accelerations** | `[2, 4, 6, 8, 10]` |
-| **Center fractions** | `[0.16, 0.08, 0.06, 0.04, 0.02]` |
+| **Accelerations (R)** | `[2, 4, 6, 8, 10]` |
+| **Center fractions (ACS)** | `[0.16, 0.08, 0.06, 0.04, 0.02]` (paired with R: 2→0.16, 4→0.08, 6→0.06, 8→0.04, 10→0.02) |
 | **Schemes** | FastMRIEquispaced, FastMRIRandom, Gaussian1D, Gaussian2D, VariableDensityPoisson, Radial |
 
-The released inference YAMLs use that same acceleration / ACS schedule.
-Default mask *name* follows the validation logging protocol:
+Released inference YAMLs keep that full R / ACS list. Default mask *name*
+matches the validation-logging protocol above.
 
-| Name | Dataset | Default mask |
-|------|---------|--------------|
-| `vsharp_brain` | fastMRI brain | FastMRIRandom |
-| `vsharp_knee` | fastMRI knee | FastMRIEquispaced |
-| `vsharp_prostate` | fastMRI prostate | FastMRIEquispaced |
-| `vsharp_cardiac` | CMRxRecon 2023 | FastMRIEquispaced |
+## Changing acceleration / scheme at inference
 
-Change `inference.dataset.transforms.masking.name` to any training scheme, or
-set a single acceleration (e.g. `accelerations: [8]`, `center_fractions: [0.04]`)
-for fixed-R evaluation.
+Edit `inference.dataset.transforms.masking` in the YAML:
 
-Weights are converted to the current DIRECT ModConv parameter layout.
+```yaml
+inference:
+  dataset:
+    transforms:
+      masking:
+        name: FastMRIEquispaced   # <- sampling scheme
+        accelerations:            # <- R factor(s)
+          - 2
+          - 4
+          - 6
+          - 8
+          - 10
+        center_fractions:         # <- ACS fraction(s), same length as accelerations
+          - 0.16
+          - 0.08
+          - 0.06
+          - 0.04
+          - 0.02
+```
+
+Examples:
+
+- Fixed **8×** with matching ACS: `accelerations: [8]`, `center_fractions: [0.04]`
+- Fixed **4×**: `accelerations: [4]`, `center_fractions: [0.08]`
+- Different scheme: set `name: FastMRIRandom` (or any training scheme above)
+
+`name`, `accelerations`, and `center_fractions` are independent — swap the
+scheme without changing R/ACS, or pin R/ACS without changing the scheme.
 
 ## Usage