modeling_summary_evaluator.py

import os, json
import numpy as np
import torch
from torch import nn
from transformers import AutoModel, AutoTokenizer

def mean_pool(last_hidden_state, attention_mask):
    mask = attention_mask.unsqueeze(-1).type_as(last_hidden_state)
    summed = (last_hidden_state * mask).sum(dim=1)
    counts = mask.sum(dim=1).clamp(min=1e-9)
    return summed / counts

def build_all_segments(ids, seg_len, overlap):
    n = len(ids)
    if n == 0: return [[]]
    if n <= seg_len: return [ids]
    step = max(1, seg_len - overlap)
    segs, start = [], 0
    while start < n:
        segs.append(ids[start:start + seg_len])
        if start + seg_len >= n: break
        start += step
    return segs

def sample_segments_cover_whole_doc(all_segs, cap):
    if len(all_segs) <= cap: return all_segs
    idx = np.linspace(0, len(all_segs)-1, cap)
    idx = np.unique(np.round(idx).astype(int)).tolist()
    idx = sorted(idx)[:cap]
    return [all_segs[i] for i in idx]

class MultiEvalSumVietN(nn.Module):
    def __init__(self, base_dir):
        super().__init__()
        with open(os.path.join(base_dir, "arch_config.json"), "r", encoding="utf-8") as f:
            cfg = json.load(f)
        self.cfg = cfg
        self.backbone = AutoModel.from_pretrained(base_dir)
        if hasattr(self.backbone.config, "use_cache"):
            self.backbone.config.use_cache = False

        hidden_in = cfg["trunk"]["hidden_in"]
        hidden_mid = cfg["trunk"]["hidden_mid"]
        dropout = cfg["trunk"]["dropout"]

        self.trunk = nn.Sequential(nn.Linear(hidden_in, hidden_mid), nn.GELU(), nn.Dropout(dropout))
        self.head_faith = nn.Linear(hidden_mid, 1)
        self.head_coh   = nn.Linear(hidden_mid, 1)
        self.head_rel   = nn.Linear(hidden_mid, 1)

        self.trunk.load_state_dict(torch.load(os.path.join(base_dir, "trunk.pt"), map_location="cpu"))
        self.head_faith.load_state_dict(torch.load(os.path.join(base_dir, "head_faith.pt"), map_location="cpu"))
        self.head_coh.load_state_dict(torch.load(os.path.join(base_dir, "head_coh.pt"), map_location="cpu"))
        self.head_rel.load_state_dict(torch.load(os.path.join(base_dir, "head_rel.pt"), map_location="cpu"))

        self.agg_type = cfg.get("agg_type", "mean")
        if self.agg_type == "attn":
            raise FileNotFoundError("agg_type='attn' requires seg_attn.pt export+upload. Set agg_type='mean' for now.")

        self.eval()

    @torch.no_grad()
    def forward(self, input_ids_3d, attention_mask_3d, seg_mask_2d):
        B, K, T = input_ids_3d.shape
        x = input_ids_3d.view(B*K, T)
        a = attention_mask_3d.view(B*K, T)

        out = self.backbone(input_ids=x, attention_mask=a).last_hidden_state
        pooled = mean_pool(out, a).view(B, K, -1)

        mask = seg_mask_2d.unsqueeze(-1).float()
        pooled = pooled * mask

        denom = mask.sum(dim=1).clamp_min(1e-6)
        doc_repr = pooled.sum(dim=1) / denom  # mean aggregation

        z = self.trunk(doc_repr)
        y = torch.cat([self.head_faith(z), self.head_coh(z), self.head_rel(z)], dim=1)
        return y

def load_for_inference(base_dir, device=None):
    tok = AutoTokenizer.from_pretrained(base_dir, use_fast=True)
    mdl = MultiEvalSumVietN(base_dir)
    if device is None:
        device = "cuda" if torch.cuda.is_available() else "cpu"
    mdl.to(device).eval()
    return mdl, tok, device

def encode_full_doc(tok, docs, sums):
    with open(os.path.join(tok.name_or_path, "arch_config.json"), "r", encoding="utf-8") as f:
        cfg = json.load(f)

    max_len = int(cfg["max_len"])
    sum_max_len = int(cfg["sum_max_len"])
    seg_len = int(cfg["seg_len"])
    seg_overlap = int(cfg["seg_overlap"])
    cap = int(cfg["max_segs_cap"])

    if tok.pad_token_id is None:
        tok.pad_token = tok.eos_token if tok.eos_token is not None else "[PAD]"
    pad_id = tok.pad_token_id

    tok_s = tok(sums, truncation=True, max_length=sum_max_len, add_special_tokens=False, return_attention_mask=False)
    sum_ids_list = tok_s["input_ids"]
    tok_d = tok(docs, truncation=False, add_special_tokens=False, return_attention_mask=False)
    doc_ids_list = tok_d["input_ids"]

    segs_per_sample = []
    for ids in doc_ids_list:
        all_segs = build_all_segments(ids, seg_len, seg_overlap)
        segs = sample_segments_cover_whole_doc(all_segs, cap)
        segs_per_sample.append(segs)

    B = len(docs)
    K = max(len(s) for s in segs_per_sample)

    flat_ids, flat_attn, flat_segmask = [], [], []
    for i in range(B):
        segs = segs_per_sample[i]
        for seg in segs:
            pair = tok.prepare_for_model(seg, sum_ids_list[i], truncation="only_first",
                                         max_length=max_len, add_special_tokens=True,
                                         return_attention_mask=True)
            flat_ids.append(torch.tensor(pair["input_ids"], dtype=torch.long))
            flat_attn.append(torch.tensor(pair["attention_mask"], dtype=torch.long))
            flat_segmask.append(1)
        for _ in range(len(segs), K):
            flat_ids.append(torch.tensor([pad_id], dtype=torch.long))
            flat_attn.append(torch.tensor([1], dtype=torch.long))
            flat_segmask.append(0)

    T = max(x.numel() for x in flat_ids)
    T = ((T + 7)//8)*8

    def pad_2d(xs):
        out = []
        for x in xs:
            if x.numel() < T:
                out.append(torch.cat([x, torch.full((T-x.numel(),), pad_id, dtype=torch.long)]))
            else:
                out.append(x)
        return torch.stack(out, dim=0)

    ids = pad_2d(flat_ids).view(B, K, T)
    attn = pad_2d(flat_attn).view(B, K, T)
    segm = torch.tensor(flat_segmask, dtype=torch.long).view(B, K)
    return ids, attn, segm