app.py

from pathlib import Path

from fastapi import FastAPI
from pydantic import BaseModel
from typing import List

import torch
from transformers import (
    AutoTokenizer,
    AutoModelForTokenClassification,
    AutoModelForSequenceClassification,
    pipeline,
)
import os


print("📁 ZAWARTOŚĆ /app/bert-aspect-ner:")
print(os.listdir("/app/bert-aspect-ner") if os.path.exists("/app/bert-aspect-ner") else "❌ NIE ISTNIEJE")

print("📁 ZAWARTOŚĆ /app/absa-roberta:")
print(os.listdir("/app/absa-roberta") if os.path.exists("/app/absa-roberta") else "❌ NIE ISTNIEJE")
# ────────────────────── konfiguracja ──────────────────────
device = "cuda" if torch.cuda.is_available() else "cpu"

ROOT = Path(__file__).parent

# Uwaga: NIE ma katalogu models/
aspect_dir = ROOT / "bert-aspect-ner"
sentiment_dir = ROOT / "absa-roberta"
# ────────────────────── modele lokalne ─────────────────────
aspect_tokenizer = AutoTokenizer.from_pretrained(
    str(aspect_dir), local_files_only=True, use_fast=False        # ← jeśli brak tokenizer.json
)
aspect_model = AutoModelForTokenClassification.from_pretrained(
    str(aspect_dir), local_files_only=True
).to(device)

sentiment_tokenizer = AutoTokenizer.from_pretrained(
    str(sentiment_dir), local_files_only=True
)
sentiment_model = AutoModelForSequenceClassification.from_pretrained(
    str(sentiment_dir), local_files_only=True
).to(device)

# ────────────────────── modele tłumaczeń (on-line) ─────────
pl_to_en = pipeline(
    "translation",
    model="Helsinki-NLP/opus-mt-pl-en",
    device=0 if device == "cuda" else -1,
)
en_to_pl = pipeline(
    "translation",
    model="gsarti/opus-mt-tc-en-pl",
    device=0 if device == "cuda" else -1,
)

# ────────────────────── schemy Pydantic ────────────────────
class Comment(BaseModel):
    text: str


class AspectSentiment(BaseModel):
    aspect: str
    sentiment: str


class AnalysisResult(BaseModel):
    results: List[AspectSentiment]

# === Słownik aliasów aspektów EN→PL (taki sam jak wcześniej) ===
aspect_aliases = {
    "food": "jedzenie", "service": "obsługa", "price": "cena",
    "taste": "smak", "waiter": "obsługa", "dish": "danie",
    "portion": "porcja", "staff": "obsługa", "decor": "wystrój",
    "menu": "menu", "drink": "napoje", "location": "lokalizacja",
    "time": "czas oczekiwania", "cleanliness": "czystość", "smell": "zapach",
    "value": "cena", "experience": "doświadczenie", "recommendation": "ogólna ocena",
    "children": "dzieci", "family": "rodzina", "pet": "zwierzęta"
}
def translate_pl_to_en(texts):
    return [res["translation_text"] for res in pl_to_en(texts)]


def translate_en_to_pl(texts):
    return [res["translation_text"] for res in en_to_pl(texts)]


def extract_aspects(text_en: str):
    inputs = aspect_tokenizer(
        text_en, return_tensors="pt", truncation=True, padding=True
    ).to(device)
    with torch.no_grad():
        outputs = aspect_model(**inputs)

    preds = torch.argmax(outputs.logits, dim=2)[0].cpu().numpy()
    tokens = aspect_tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
    labels = [aspect_model.config.id2label[p] for p in preds]

    aspects, current_tokens = [], []
    for token, label in zip(tokens, labels):
        if label == "B-ASP":
            if current_tokens:
                aspects.append(aspect_tokenizer.convert_tokens_to_string(current_tokens).strip())
            current_tokens = [token]
        elif label == "I-ASP" and current_tokens:
            current_tokens.append(token)
        else:
            if current_tokens:
                aspects.append(aspect_tokenizer.convert_tokens_to_string(current_tokens).strip())
                current_tokens = []
    if current_tokens:
        aspects.append(aspect_tokenizer.convert_tokens_to_string(current_tokens).strip())

    # ↓ usuń spacje z „##” i zduplikowane wyniki
    return list({tok.replace(" ##", "") for tok in aspects})


# ────────────────────── FastAPI ────────────────────────────
app = FastAPI()


@app.post("/analyze", response_model=AnalysisResult)
def analyze_comment(comment: Comment):
    text_pl = comment.text
    text_en = translate_pl_to_en([text_pl])[0]
    aspects = extract_aspects(text_en)

    results: list[AspectSentiment] = []
    for asp in aspects:
        input_text = f"{text_en} [SEP] {asp}"
        inputs = sentiment_tokenizer(
            input_text, return_tensors="pt", truncation=True, padding=True
        ).to(device)
        with torch.no_grad():
            logits = sentiment_model(**inputs).logits
            predicted_class_id = int(logits.argmax().cpu())
            sentiment_label = {
                0: "negatywny",
                1: "neutralny",
                2: "pozytywny",
                3: "konfliktowy",
            }[predicted_class_id]

        asp_pl = aspect_aliases.get(asp, translate_en_to_pl([asp])[0].lower())
        results.append(AspectSentiment(aspect=asp_pl, sentiment=sentiment_label))

    return {"results": results}