File size: 5,249 Bytes
c1e97d9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
# Implementation and reproducibility notes

## Effective specification

Phase 2.0 plus Phase 2.1 Amendment A01 is the specification. Runtime code uses typed Grid/Condition/Bem/Recipe objects and an explicit job ledger. The protocol files are shipped for audit; arbitrary JSON keys are not silently interpreted as runtime hyperparameter overrides. Changing a frozen value requires editing the typed recipe and recording a numbered amendment before confirmatory output inspection. `study.json` hashes code, protocol documents and pinned fidelity sources/fixtures.

The package's semantic version is 3.0.0; the scientific protocol remains 2.1/A01. Tiny smoke runs are explicitly separated. Main training records the actual selected LR, epoch, checkpoint and initialization hashes. Test summaries link to data manifests and checkpoints. Source development attempts remain in `runs/TUNE_*`.

## Fidelity

FortiTran and AdaFortiTran are independent implementations compared to the pinned released operators with identical weights and inputs. Shared RI weights, layer cloning, 512-row positional storage, metadata dimensions/units and patch ordering are retained. Their labels remain released-code variants.

HELENA uses the original saved native-grid Keras model to validate the port; only the 120-carrier reshape/token adaptation is used for new common-data training. Native pretrained weights are a fidelity fixture, not initialization for the new study. The main model is initialized from scratch. See `verification/` for the delivered numerical check record.

Channelformer is checked against an independent NumPy implementation of the MATLAB-defined selected operator sequence. This establishes those operators' agreement; it does not establish unavailable native MATLAB checkpoint parity. ReEsNet uses the audited supplied-project adaptation. Both qualifications must remain in the paper.

All redistributed author material retains its license in `vendor/`. The native HELENA binary was retrieved freshly from the pinned commit to obtain valid bytes; its delivered SHA-256 is recorded in the package manifest. Old scratch/source-lock records do not replace this byte-level verification.

## Precision and device behavior

Channel path generation and classical solves use float64/complex128. Labels/pilot tensors are stored as complex64 at the neural boundary. Metrics convert operands to complex128 before subtracting and use float64 reductions. Model loss is the sum of real/imaginary squared error per complex RE, averaged over the grid; the R backward loss alone retains the released one-half scaling.

Device-resident source data are a performance optimization. They preserve source examples, explicit shuffle permutations and final partial batches; a test compares staged and streaming tensors exactly. Generation device is recorded in the bank specification. Do not regenerate part of a bank on another device and assume bitwise-identical simulator RNG streams.

Strict determinism is requested through torch and CuBLAS configuration. CUDA acceptance includes forward/backward checks for every architecture on the actual target device. Explicit replication padding replaces a potentially nondeterministic backward kernel without changing mathematical outputs, trainable state layout or initialization.

The main FP32 GPU environment has not been benchmarked on the user's hardware during preparation. Deterministic kernels, driver compatibility, full-size memory use and long-run stability must pass locally. No claimed speedup is inferred solely from parameter/MAC counts.

## Checkpoints and restart

Main training restarts from the last completed epoch, with optimizer and RNG states restored. Checkpoint selection uses source VAL; development uses source TUNE. Exact ties keep the earlier checkpoint. A numerical divergence fails loudly. No seed is replaced. Hardware interruption may repeat the unfinished epoch.

Completed prediction conditions and completed compatible training jobs are reused. Short adaptation can be restarted deterministically from its source checkpoint. Atomic file replacement prevents partially written individual records from becoming final records. Backups may be taken between stages; provider-level synchronization is needed if you want durability during a long stage.

Do not launch concurrent writers to the same run. Do not edit source after creating a confirmatory root. Archive an invalidated root with its reason and create a new root for corrected execution. Do not manually remove a completed marker to mix versions.

## Interpretation boundaries

The scalar APP receiver uses assumed channel estimates and AWGN variance and deliberately does not receive test-derived estimation/ICI variance. Perfect diagonal and full-G references have different information access. Fractional-delay WAVE is exact for the within-CP periodically extended symbol; it does not simulate CP-violating ISI or a full NR link.

The process is designed to fail if required evidence or compatible files are missing. Passing small software checks is necessary but does not establish the scientific result. Full confidence intervals, hypothesis tests and performance conclusions require the complete planned experimental output.