Final Stage 1 (ResNet-18 Barlow Twins encoder) and Stage 2 (channel-aggregation Transformer) checkpoints used to produce all results in the manuscript.
stage1_encoder_final.pt— Stage 1 spatial encoder (single-channel ResNet-18, Barlow Twins pretraining).stage2_aggregator_final.pt— Stage 2 permutation-invariant channel Transformer (trained on Stage 1 CLS embeddings).
Load with the --checkpoint argument of code/extract_stage1_embeddings.py / code/extract_stage2_embeddings.py (see README) to skip training and go straight to embedding extraction.