web_app/model.py

"""
CEFR Sentence Level Assessment Model
Loads and runs inference with the metric proto k3 model
"""

import re
from pathlib import Path
from typing import List, Tuple, Dict

import torch
from transformers import AutoTokenizer, AutoModel


class PrototypeClassifier(torch.nn.Module):
    """Metric-based prototype classifier for CEFR level assessment"""

    def __init__(
        self,
        encoder,
        num_labels: int,
        hidden_size: int,
        prototypes_per_class: int,
        temperature: float = 10.0,
        layer_index: int = -2,
    ):
        super().__init__()
        self.encoder = encoder
        self.num_labels = num_labels
        self.prototypes_per_class = prototypes_per_class
        self.temperature = temperature
        self.layer_index = layer_index
        self.prototypes = torch.nn.Parameter(
            torch.empty(num_labels, prototypes_per_class, hidden_size)
        )

    def set_prototypes(self, proto_tensor: torch.Tensor) -> None:
        """Set prototype weights"""
        with torch.no_grad():
            self.prototypes.copy_(proto_tensor)

    def encode(self, input_ids, attention_mask, token_type_ids=None) -> torch.Tensor:
        """Encode input sentences to normalized embeddings"""
        outputs = self.encoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            output_hidden_states=True,
        )
        hidden = outputs.hidden_states[self.layer_index]
        # mean pooling
        mask = attention_mask.unsqueeze(-1).float()
        summed = torch.sum(hidden * mask, dim=1)
        counts = torch.clamp(mask.sum(dim=1), min=1e-9)
        pooled = summed / counts
        pooled = torch.nn.functional.normalize(pooled, p=2, dim=1)
        return pooled

    def forward(self, input_ids, attention_mask, token_type_ids=None):
        """Forward pass returning logits"""
        x = self.encode(input_ids, attention_mask, token_type_ids)
        # cosine similarity with prototypes, average over K for each class
        protos = torch.nn.functional.normalize(self.prototypes, p=2, dim=-1)
        # [B, H] x [C,K,H] -> [B,C,K]
        sim = torch.einsum("bh,ckh->bck", x, protos)
        sim_mean = sim.mean(dim=2)  # average over K
        logits = sim_mean * self.temperature
        return {"logits": logits}

    def predict(self, input_ids, attention_mask, token_type_ids=None) -> torch.Tensor:
        """Predict CEFR levels"""
        outputs = self.forward(input_ids, attention_mask, token_type_ids)
        return torch.argmax(outputs["logits"], dim=1)


class CEFRModel:
    """Wrapper class for CEFR assessment model"""

    def __init__(self, model_path: str = None, device: str = None):
        """
        Initialize the CEFR assessment model

        Args:
            model_path: Path to the trained model checkpoint
            device: Device to run inference on ('cuda' or 'cpu')
        """
        if device is None:
            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        else:
            self.device = torch.device(device)

        # CEFR level mapping
        self.id_to_label = {0: "A1", 1: "A2", 2: "B1", 3: "B2", 4: "C1", 5: "C2"}
        self.label_to_id = {v: k for k, v in self.id_to_label.items()}

        # Model parameters
        self.model_name = "KB/bert-base-swedish-cased"
        self.hidden_size = 768
        self.num_labels = 6
        self.prototypes_per_class = 3
        self.temperature = 10.0

        # Load tokenizer
        self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)

        # Load model
        encoder = AutoModel.from_pretrained(self.model_name)
        self.model = PrototypeClassifier(
            encoder=encoder,
            num_labels=self.num_labels,
            hidden_size=self.hidden_size,
            prototypes_per_class=self.prototypes_per_class,
            temperature=self.temperature,
        )

        # Load trained weights
        if model_path is None:
            # Try to find the model automatically
            default_paths = [
                "runs/metric-proto-k3/metric_proto.pt",
                "runs/metric-proto/metric_proto.pt",
                "runs/bert-baseline/bert_baseline.pt",
                "../runs/metric-proto-k3/metric_proto.pt",  # Relative to web_app/
            ]
            for path in default_paths:
                if Path(path).exists():
                    model_path = path
                    print(f"Auto-detected model: {model_path}")
                    break

        if model_path:
            # Try different relative paths
            possible_paths = [
                Path(model_path),
                Path(__file__).parent / model_path,
                Path(__file__).parent.parent / model_path,
            ]

            checkpoint = None
            for path in possible_paths:
                if path.exists():
                    print(f"Loading model from {path}")
                    checkpoint = torch.load(path, map_location=self.device, weights_only=False)
                    break

            if checkpoint is None:
                print(f"Warning: Model file not found at {model_path}")
                print("Model will be initialized with random weights!")
        else:
            print("Warning: No model path specified. Model will be initialized with random weights!")
            checkpoint = None

        if checkpoint is not None:

            # Load model state dict
            if "state_dict" in checkpoint:
                state_dict = checkpoint["state_dict"]
                # Handle DataParallel state dict
                new_state_dict = {}
                for key, value in state_dict.items():
                    if key.startswith("model."):
                        new_key = key[6:]  # Remove 'model.' prefix
                    else:
                        new_key = key
                    new_state_dict[new_key] = value
                self.model.load_state_dict(new_state_dict, strict=False)
            else:
                self.model.load_state_dict(checkpoint)

            # Load prototypes if available
            if "prototypes" in checkpoint:
                self.model.set_prototypes(checkpoint["prototypes"].to(self.device))

        self.model.to(self.device)
        self.model.eval()

    def tokenize(self, texts: List[str], max_length: int = 128) -> Dict[str, torch.Tensor]:
        """Tokenize input texts"""
        encoded = self.tokenizer(
            texts,
            truncation=True,
            padding=True,
            max_length=max_length,
            return_tensors="pt",
        )
        return encoded

    def predict_batch(self, sentences: List[str]) -> List[Tuple[str, float]]:
        """
        Predict CEFR levels for a batch of sentences

        Args:
            sentences: List of sentences to assess

        Returns:
            List of (level, confidence) tuples
        """
        if not sentences:
            return []

        # Tokenize
        encoded = self.tokenize(sentences)
        input_ids = encoded["input_ids"].to(self.device)
        attention_mask = encoded["attention_mask"].to(self.device)

        # Predict
        with torch.no_grad():
            logits = self.model(input_ids, attention_mask)["logits"]
            probs = torch.softmax(logits, dim=1)
            predictions = torch.argmax(logits, dim=1)

        # Format results
        results = []
        cpu_probs = probs.cpu()
        for i, pred in enumerate(predictions.cpu().numpy()):
            level = self.id_to_label[pred]
            confidence = float(cpu_probs[i][pred].item())
            # Handle NaN values
            if torch.isnan(cpu_probs[i][pred]):
                confidence = 1.0 / self.num_labels
            results.append((level, confidence))

        return results

    def predict_sentence(self, sentence: str) -> Tuple[str, float]:
        """Predict CEFR level for a single sentence"""
        results = self.predict_batch([sentence])
        return results[0]


def split_into_sentences(text: str) -> List[str]:
    """
    Split text into sentences

    Args:
        text: Input text (Swedish)

    Returns:
        List of sentences
    """
    # Simple sentence splitting based on punctuation
    # Swedish sentence endings: . ! ?
    # Split on punctuation followed by space and uppercase letter, or end of string

    sentences = re.split(r'([.!?])\s+', text)

    # Combine punctuation with previous sentence
    combined = []
    for i in range(0, len(sentences) - 1, 2):
        if i + 1 < len(sentences):
            combined.append(sentences[i] + sentences[i + 1])
        else:
            combined.append(sentences[i])

    # Handle the last sentence if there's no punctuation
    if len(sentences) % 2 == 1 and sentences[-1].strip():
        combined.append(sentences[-1])

    # Clean up sentences
    cleaned = []
    for sent in combined:
        sent = sent.strip()
        if sent:
            cleaned.append(sent)

    return cleaned


def assess_text(text: str, model: CEFRModel) -> List[Dict[str, any]]:
    """
    Assess a text and return sentence-level CEFR annotations

    Args:
        text: Input text (Swedish)
        model: CEFR assessment model

    Returns:
        List of dictionaries with sentence and level information
    """
    # Split text into sentences
    sentences = split_into_sentences(text)

    if not sentences:
        return []

    # Predict CEFR levels
    predictions = model.predict_batch(sentences)

    # Format results
    results = []
    for sent, (level, confidence) in zip(sentences, predictions):
        results.append({
            "sentence": sent,
            "level": level,
            "confidence": confidence,
        })

    return results