seismic-benchmark-code

Release Highlights

This release focuses on evaluation reliability, visualization consistency, and training reproducibility across all four seismic restoration tasks:

• Interpolation
• Paired interpolation
• Coherent noise attenuation
• Residual coherent noise attenuation

---

1. Unified Visualization Scaling

All training and inference visualizations now use a shared symmetric color scale by default.

Previously, residual panels used independent adaptive scaling, which visually exaggerated small residual amplitudes and made direct comparison difficult. This release introduces:

• Global vmin/vmax support in visualization utilities
• Shared scaling across input / prediction / target / residual panels
• Run-level fixed color scaling during inference
• Optional fallback to adaptive per-panel scaling (share_scale=False)

Improvements

• Residual amplitudes are now physically comparable to reconstructed signals
• Figures from different shots within the same inference run are visually consistent
• Training and inference outputs are easier to interpret and analyze quantitatively

---

2. Shot-Level (FFID-Based) Dataset Splitting

A new shot-level train/validation/test splitting pipeline has been added to eliminate data leakage caused by patch-level random splitting.

The new workflow:

• Splits data by unique FFID (shot number)
• Applies deterministic sequential partitioning before patch extraction
• Patchifies each subset independently
• Supports distributed training and optional test-shot export

This feature is integrated into all four training pipelines and ensures:

• No overlap between train/val/test patches from the same shot gather
• Reproducible and deterministic dataset partitions
• More reliable validation and benchmark evaluation

Backward compatibility is preserved: existing configurations without shot_split continue using the legacy random patch split.

---

3. Automatic Best-Validation Checkpoint Saving

Training now automatically maintains a best.pt checkpoint corresponding to the lowest validation loss observed during training.

New functionality includes:

• Validation-aware checkpoint tracking
• Automatic overwrite when validation performance improves
• Preservation of model, optimizer, scheduler, and epoch states

This is now enabled in all four training scripts while keeping periodic epoch checkpoints unchanged.

Benefits

• Easier model selection after long training runs
• Safer recovery of optimal parameters
• Simplified downstream inference and evaluation workflows

while maintaining backward compatibility with existing training and inference configurations.