dmusingu/lapvqa-rrg-native

LAPVQA — Radiology Report Generation (Native / End-to-end)

Part of the LAPVQA collection.

Description

RRG decoders trained end-to-end alongside their vision encoders. Each checkpoint is a dict: {state_dict, vis_dim, d_model, num_layers, nhead, encoder, epoch, val_bleu4}.

File	Encoder	vis_dim
clip-vit-l14.pt	CLIP ViT-L/14 (fine-tuned)	1024
siglip.pt	SigLIP (fine-tuned)	1152
florence2.pt	Florence-2 (fine-tuned)	1024
coca.pt	CoCa (fine-tuned)	768
mae-vit-l16.pt	MAE ViT-L/16 (fine-tuned)	1024

Results (MIMIC-CXR test set, MAE-ViT-L/16)

BLEU-4	ROUGE-L	RadGraph-s
0.032	0.164	0.195

Loading

import torch
from lapvqa.rrg.heads import ReportGenerationHead

ckpt = torch.load("mae-vit-l16.pt", map_location="cpu")
head = ReportGenerationHead(
    vis_dim    = ckpt["vis_dim"],
    d_model    = ckpt["d_model"],
    num_layers = ckpt["num_layers"],
    nhead      = ckpt["nhead"],
)
head.load_state_dict(ckpt["state_dict"])
head.eval()