dmusingu / lapvqa-rrg

LAPVQA — Radiology Report Generation (Frozen Off-the-shelf Encoders)

Part of the LAPVQA collection .

Description

Autoregressive decoder heads for Radiology Report Generation on MIMIC-CXR, trained on top of five frozen off-the-shelf encoders. Each checkpoint is a dict: {state_dict, vis_dim, d_model, num_layers, nhead, encoder, epoch, val_bleu4} .

Architecture — ReportGenerationHead

vis_proj : Linear(vis_dim → 512)
tok_emb  : Embedding(50257, 512)   # GPT-2 vocab, weight-tied with lm_head
pos_emb  : Embedding(150, 512)
decoder  : 6 × TransformerDecoderLayer (pre-norm)
lm_head  : Linear(512 → 50257, bias=False)

Results (MIMIC-CXR test set)

Loading

import torch
import tiktoken
from lapvqa.rrg.heads import ReportGenerationHead

ckpt = torch.load("siglip.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()

enc = tiktoken.get_encoding("gpt2")
bos_id = eos_id = enc.eot_token

# vis_tokens: [B, N, vis_dim] — patch tokens from the frozen encoder
token_ids = head.generate(vis_tokens, bos_id=bos_id, eos_id=eos_id, max_len=150)
reports   = [enc.decode(ids) for ids in token_ids]