PG-XFormer (MIUA 2026)

This repository contains the pre-trained and fine-tuned weights for PG-XFormer.

Model Description

• Architecture: VideoMAE + CTR-GCN with Bottleneck Cross-Attention Fusion.
• Stage A Weights: Synthetically pre-trained on InfiniteRep.
• Stage B Weights: Fine-tuned on the real-world KIMORE/UI-PRMD clinical datasets using LoRA and MMD.

Included Files

• pgxformer_synthetic_best.pt: Base checkpoint from Stage A.
• pgxformer_finetuned_best.pt: Adapted checkpoint from Stage B.