app.py

import os
import re
import json
import torch
import gradio as gr
from transformers import AutoTokenizer, AutoModel
from huggingface_hub import hf_hub_download

MODEL_ID = "usmannawaz/ocs"

MAX_LENGTH = 512
STRIDE = 0
MAX_SEGS_EVAL = 128
SEG_FORWARD_BS = 8

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
os.environ["TOKENIZERS_PARALLELISM"] = "false"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)
encoder = AutoModel.from_pretrained(MODEL_ID).to(DEVICE)
encoder.eval()
hidden = encoder.config.hidden_size

label_path = hf_hub_download(MODEL_ID, "label2id.json")
with open(label_path, "r", encoding="utf-8") as f:
    label2id = json.load(f)
id2label = {v: k for k, v in label2id.items()}

head_path = hf_hub_download(MODEL_ID, "doc_classifier_head.pt")
head_state = torch.load(head_path, map_location="cpu")

classifier = torch.nn.Linear(hidden, len(label2id))
classifier.load_state_dict(head_state)
classifier = classifier.to(DEVICE)
classifier.eval()

LABEL_INFO = {
    "Old Church Slavonic": {"century": "9th–11th", "language": "Old Church Slavonic"},
    "Church Slavonic": {"century": "12th–17th", "language": "Church Slavonic"},
    "New Church Slavonic": {"century": "18th", "language": "New Church Slavonic"},
    "Ruthenian": {"century": "15th–18th", "language": "Ruthenian"},
}

BIBLE_REF_RE = re.compile(
    r"""\b(?:[^\W\d_]{1,20}\.?(?:\s+[^\W\d_]{1,20}\.?)?)\s*\d{1,3}\s*[:.,]\s*\d{1,3}(?:\s*[-–]\s*\d{1,3})?\b""",
    re.UNICODE,
)
NOISY_SYMBOLS_RE = re.compile(r"[†‡§¶¦|‖•·◦※◆◇■□▲△▼▽★☆✝︎✟✠]+", re.UNICODE)

def preprocess_text(t: str) -> str:
    if not t:
        return ""
    t = BIBLE_REF_RE.sub(" ", t)
    t = re.sub(r"\b\d+[A-Za-zА-Яа-я]{1,3}\b", " ", t)
    t = re.sub(r"\b\d{1,3}\b", " ", t)
    t = re.sub(r"\[\s*\.{1,}\s*\]", " ", t)
    t = re.sub(r"\[([^\W\d_]{1,3})\]", r"\1", t, flags=re.UNICODE)
    t = NOISY_SYMBOLS_RE.sub(" ", t)
    t = re.sub(r"\s+", " ", t).strip()
    return t.lower()

@torch.no_grad()
def predict(text: str) -> str:
    if not text or not text.strip():
        return ""

    text = preprocess_text(str(text))
    if not text.strip():
        return ""

    tok = tokenizer(
        text,
        truncation=True,
        max_length=MAX_LENGTH,
        padding="max_length",
        return_overflowing_tokens=True,
        stride=STRIDE,
        return_tensors=None,
    )

    input_ids = tok["input_ids"]
    attention_mask = tok["attention_mask"]

    if MAX_SEGS_EVAL is not None and len(input_ids) > MAX_SEGS_EVAL:
        input_ids = input_ids[:MAX_SEGS_EVAL]
        attention_mask = attention_mask[:MAX_SEGS_EVAL]

    S = len(input_ids)
    input_ids = torch.tensor(input_ids, dtype=torch.long, device=DEVICE)
    attention_mask = torch.tensor(attention_mask, dtype=torch.long, device=DEVICE)

    sum_emb = torch.zeros((hidden,), device=DEVICE, dtype=torch.float32)
    count = 0

    for start in range(0, S, SEG_FORWARD_BS):
        end = min(S, start + SEG_FORWARD_BS)
        out = encoder(
            input_ids=input_ids[start:end, :],
            attention_mask=attention_mask[start:end, :],
        )
        seg_emb = out.last_hidden_state[:, 0, :]
        sum_emb += seg_emb.float().sum(dim=0)
        count += seg_emb.shape[0]

    doc_emb = sum_emb / max(count, 1)
    logits = classifier(doc_emb.unsqueeze(0))[0]
    pred_id = int(torch.argmax(logits).item())
    label = id2label[pred_id]

    info = LABEL_INFO.get(label)
    if not info:
        return "—"

    return f"Language: {info['language']}\nCentury: {info['century']}"

demo = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(lines=12, placeholder="Paste a document here..."),
    outputs=gr.Textbox(label="Century & Language"),
    title="Document Classifier",
    description="Paste a document and get the predicted historical range.",
)

demo.queue()
demo.launch(server_name="0.0.0.0", server_port=7860, ssr_mode=False)