src/classifier.py

# newQlasifier.py
from __future__ import annotations
from pathlib import Path
from functools import lru_cache
import numpy as np
import pickle
import os

os.environ.setdefault("SENTENCE_TRANSFORMERS_HOME", "/app/.cache/sentence_transformers")
os.environ.setdefault("HF_HOME", "/app/.cache/huggingface")
os.environ.setdefault("HUGGINGFACE_HUB_CACHE", "/app/.cache/huggingface")
os.environ.setdefault("TRANSFORMERS_CACHE", "/app/.cache/huggingface/transformers")
os.environ.setdefault("TRANSFORMERS_NO_TF", "1")
os.environ.setdefault("TRANSFORMERS_NO_FLAX", "1")


# ---- Stałe i ścieżki (ABSOLUTNE względem tego pliku) ----
THIS_DIR = Path(__file__).resolve().parent
MODEL_A_PATH = THIS_DIR / "best_model_70%.keras"
MODEL_B_PATH = THIS_DIR / "best_model_70%1.keras"
MLB_A_PATH   = THIS_DIR / "mlb.pkl"
MLB_B_PATH   = THIS_DIR / "mlb1.pkl"
EMBED_NAME   = "paraphrase-multilingual-MiniLM-L12-v2"   # 384-D

# ---- Ładowanie zależności ciężkich (lazy + cache) ----
from sentence_transformers import SentenceTransformer
EMBED_NAME = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"  # Twój model

@lru_cache(maxsize=1)
def _embedder():
    cache_dir = os.getenv("SENTENCE_TRANSFORMERS_HOME", "/app/.cache/sentence_transformers")
    return SentenceTransformer(EMBED_NAME, cache_folder=cache_dir)


def _load_with_fallback(model_path: Path):
    """
    Najpierw spróbuj tf.keras, a jeśli trafi się konflikt deserializacji (np. 'batch_shape'),
    spróbuj standalone 'keras'. Dzięki temu działa w różnych środowiskach.
    """
    # 1) tf.keras
    try:
        import tensorflow as tf
        return tf.keras.models.load_model(str(model_path), compile=False)
    except TypeError as e:
        # typowy błąd z 'batch_shape' przy niezgodnych wersjach
        err = str(e).lower()
        if "unrecognized keyword arguments" in err or "batch_shape" in err:
            pass  # spróbujemy standalone keras
        else:
            raise
    except Exception:
        # inne problemy też spróbujmy obejść via keras
        pass

    # 2) standalone keras
    import keras
    return keras.models.load_model(str(model_path), compile=False)

@lru_cache(maxsize=1)
def _model_a():
    return _load_with_fallback(MODEL_A_PATH)

@lru_cache(maxsize=1)
def _model_b():
    return _load_with_fallback(MODEL_B_PATH)

@lru_cache(maxsize=1)
def _mlb_a():
    with open(MLB_A_PATH, "rb") as f:
        return pickle.load(f)

@lru_cache(maxsize=1)
def _mlb_b():
    with open(MLB_B_PATH, "rb") as f:
        return pickle.load(f)

# ---- API: funkcje do wywoływania z innych plików ----
def encode_text(text: str) -> np.ndarray:
    """
    Zwraca wektor (1, d) jako float32.
    """
    emb = _embedder()
    X = emb.encode([text], convert_to_numpy=True, show_progress_bar=False)
    return np.asarray(X, dtype="float32")

def predict_raw(text: str) -> str:
    """
    Predykcja modelem A (best_model_70%.keras) -> zwraca etykietę (string).
    """
    X = encode_text(text)           # (1, d)
    y = _model_a().predict(X, verbose=0)[0]  # (n_classes,)
    cls = int(np.argmax(y))
    return _mlb_a().classes_[cls]

def predict_raw1(text: str) -> str:
    """
    Predykcja modelem B (best_model_70%1.keras) -> zwraca etykietę (string).
    """
    X = encode_text(text)
    y = _model_b().predict(X, verbose=0)[0]
    cls = int(np.argmax(y))
    return _mlb_b().classes_[cls]