model.py

"""Shared MedLayEval model definition.

A LoRA-adapted Qwen2.5-VL backbone with a 5-attribute regression head.
Used by:
  - train.py     (distillation training)
  - validate.py  (held-out correlation report)
  - inference.py (scoring benchmark VLM outputs)

The five attributes, in fixed order, are:
  modality, anatomy, finding, factual, readability.
"""

import torch
import torch.nn as nn

ATTRS = ["modality", "anatomy", "finding", "factual", "readability"]


class VLMRegressor(nn.Module):
    """LoRA-adapted VLM with an attention-mask-pooled regression head.

    Forward returns shape (B, 5) sigmoid scores in [0, 1].
    """

    def __init__(self, vlm, hidden_size: int):
        super().__init__()
        self.vlm = vlm
        self.head = nn.Sequential(
            nn.Linear(hidden_size, 256),
            nn.GELU(),
            nn.Dropout(0.1),
            nn.Linear(256, 5),
            nn.Sigmoid(),
        )

    def forward(self, target_scores=None, **inputs):
        # `target_scores` is ignored at inference; accepting it lets the
        # train loop pass labels through the same forward signature.
        out = self.vlm(**inputs, output_hidden_states=True, return_dict=True)
        h = out.hidden_states[-1]                          # (B, T, H)
        mask = inputs["attention_mask"].unsqueeze(-1).to(h.dtype)
        pooled = (h * mask).sum(1) / mask.sum(1).clamp(min=1)
        return self.head(pooled.float())