Restore consistent NMR peak-set encoding under batch padding
You are repairing a scientific machine-learning workflow in the supplied NMRTrans source snapshot. The workflow encodes 1H and 13C NMR peak sets into continuous representations that a decoder consumes for molecular structure elucidation. The required representation treats spectra as unordered peak sets and uses permutation-equivariant attention blocks followed by permutation-invariant pooling.
Different compounds produce different numbers of observed peaks, so spectra are batched by appending dummy padding rows up to a common length and recording which rows are real observations. A user running this encoding workflow reports that the encoded spectrum is not stable under this batching:
- Encoding the same spectrum with a different number of appended padding rows changes the produced representations, even though the observed peaks are identical.
- In a padded batch, the pooled spectrum-level representation barely responds when a real observed peak is moved to a different chemical shift.
Both observations contradict the required set-encoding contract: the encoded representation must be a function of the observed peaks alone, and it must respond to chemically meaningful changes in those peaks.
Inspect the complete source snapshot and run python reproduce.py. Repair the implementation so that the peak-set
encoders satisfy the documented contract for all valid inputs. The repair
must generalize to arbitrary spectra, peak counts, padding lengths, random
initializations, and encoder configurations; do not hard-code the public
fixtures, specific tensor values, or a particular padding length.
Keep the overall architecture (the stacked ISAB encoder, the PMA pooling, the learnable tokens, and the projection layers) and the existing function signatures intact, and preserve the documented input feature layout for both modalities. The public reproduction is only a small smoke diagnostic; use the source call graph to determine the complete behavior that needs to be repaired. Do not use the network or add external data files. Keep changes limited to the scientific implementation and any focused comments that are necessary for a general fix.