inference.py

import numpy as np
import joblib
import torch
import torch.nn as nn
import torch.nn.functional as F
import pytorch_lightning as pl 
from sklearn.metrics.pairwise import cosine_distances
from sentence_transformers import SentenceTransformer

class IntentClassifier(pl.LightningModule):
    def __init__(self, input_dim=384, hidden_dim=256, output_dim=150, lr=1e-3, weight_decay=1e-4):
        super().__init__()
        self.save_hyperparameters()

        self.fc1 = nn.Linear(input_dim, hidden_dim)
        self.bn1 = nn.BatchNorm1d(hidden_dim)         
        self.relu = nn.ReLU()
        self.dropout = nn.Dropout(0.3)               
        self.fc2 = nn.Linear(hidden_dim, output_dim)

    def forward(self, x):
        x = self.fc1(x)
        x = self.bn1(x)           
        x = self.relu(x)
        x = self.dropout(x)        
        return self.fc2(x)
    
    def training_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        loss = F.cross_entropy(logits, y)
        self.log("train_loss", loss)
        return loss


    def validation_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        preds = torch.argmax(logits, dim=1)

    
        mask = y != -1
    
        if mask.sum() > 0:
            val_loss = F.cross_entropy(logits[mask], y[mask])
            val_acc = (preds[mask] == y[mask]).float().mean()
        else:
            val_loss = torch.tensor(0.0, device=self.device)
            val_acc = torch.tensor(0.0, device=self.device)
    
        self.log("val_loss", val_loss, prog_bar=True)
        self.log("val_acc", val_acc, prog_bar=True)

    
    def configure_optimizers(self):
        return torch.optim.Adam(
            self.parameters(), 
            lr=self.hparams.lr,
            weight_decay=1e-4  
        )

class IntentClassifierWithOOS:
    def __init__(self, embedder, classifier, oos_detector, label_encoder, centroids_dict, oos_threshold=0.5, device="cpu"):
        self.embedder = embedder  # SentenceTransformer
        self.classifier = classifier.eval().to(device)  # MLP
        self.oos_detector = oos_detector  # pipeline sklearn
        self.label_encoder = label_encoder  # fitted LabelEncoder
        self.centroids_dict = centroids_dict  # {class_id: centroid}
        self.threshold = oos_threshold
        self.device = device

    def _compute_features(self, embedding, logits, predicted_class):
        probs = F.softmax(logits, dim=0).cpu().numpy()
        entropy = -np.sum(probs * np.log(probs + 1e-10))
        msp = np.max(probs)
        energy = torch.logsumexp(logits, dim=0).item()

        # Logit gap
        sorted_logits = torch.sort(logits, descending=True).values
        logit_gap = (sorted_logits[0] - sorted_logits[1]).item()

        # Euclidean distance to class centroid
        centroid = self.centroids_dict.get(predicted_class)
        dist = np.linalg.norm(embedding - centroid) if centroid is not None else np.nan

        # Cosine distance
        cos_dist = cosine_distances([embedding], [centroid])[0][0] if centroid is not None else np.nan

        norm_emb = np.linalg.norm(embedding)

        return np.array([entropy, msp, dist])  


    def predict(self, sentence):
        # 1. Embedding
        embedding = self.embedder.encode(sentence)
        embedding = np.array(embedding)  
        embedding_tensor = torch.tensor(embedding, dtype=torch.float32).unsqueeze(0).to(self.device)

        # 2. Intent prediction (MLP)
        with torch.no_grad():
            logits = self.classifier(embedding_tensor)
            logits = logits.squeeze(0)  
        probs = F.softmax(logits, dim=0)
        predicted_class = torch.argmax(probs).item()
        confidence = probs[predicted_class].item()

        # 3. Feature extraction
        features = self._compute_features(embedding, logits, predicted_class).reshape(1, -1)

        # 4. OOS detection
        oos_score = self.oos_detector.predict_proba(features)[0, 1]
        is_oos = oos_score >= self.threshold

        # 5. Output
        return {
            "intent": "oos" if is_oos else self.label_encoder.inverse_transform([predicted_class])[0],
            "is_oos": bool(is_oos),
            "confidence": None if is_oos else confidence,
            "oos_score": oos_score
        }


# Load all saved components from the current directory
best_model = IntentClassifier.load_from_checkpoint(
    "intent_classifier.ckpt",
    map_location=torch.device("cpu")
)

oos_detector = joblib.load("oos_detector.pkl")
label_encoder = joblib.load("label_encoder.pkl")
class_centroids = joblib.load("class_centroids.pkl")
best_threshold = joblib.load("oos_threshold.pkl")

print("Model charging")
# Recharger l'embedding model
embedder = SentenceTransformer("intfloat/e5-small-v2")

# Build the full inference model
model = IntentClassifierWithOOS(
    embedder=embedder,
    classifier=best_model,
    oos_detector=oos_detector,
    label_encoder=label_encoder,
    centroids_dict=class_centroids,
    oos_threshold=best_threshold,
    device="cpu"
)