models.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import (
    AutoModel, AutoConfig, AutoTokenizer, 
    T5ForConditionalGeneration, T5Config,
    AutoModelForSequenceClassification,
    PreTrainedModel, PretrainedConfig
)
from transformers.modeling_utils import (
    load_state_dict,
    WEIGHTS_NAME,
    SAFE_WEIGHTS_NAME,
    SAFE_WEIGHTS_INDEX_NAME,
    WEIGHTS_INDEX_NAME
)
from transformers.utils import (
    is_safetensors_available,
    is_torch_available,
    logging,
    EntryNotFoundError,
    PushToHubMixin
)
import os
import json
import numpy as np

logger = logging.get_logger(__name__)

class BaseHateSpeechModel(nn.Module):
    """Base class cho tất cả các mô hình hate speech detection"""
    def __init__(self, model_name: str, num_labels: int = 3):
        super().__init__()
        self.num_labels = num_labels
        self.model_name = model_name
        
    def forward(self, input_ids, attention_mask, labels=None):
        raise NotImplementedError
    
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
        """
        Load model từ pretrained checkpoint.
        Transformers sẽ tự động load state_dict sau khi khởi tạo model.
        """
        # Extract config từ kwargs (transformers sẽ pass config vào đây)
        config = kwargs.pop("config", None)
        
        # Load config nếu chưa có
        if config is None:
            try:
                config = AutoConfig.from_pretrained(pretrained_model_name_or_path)
            except Exception:
                config = {}
        
        # Get num_labels từ config hoặc kwargs
        num_labels = kwargs.pop("num_labels", None)
        if num_labels is None:
            if hasattr(config, "num_labels"):
                num_labels = config.num_labels
            elif isinstance(config, dict) and "num_labels" in config:
                num_labels = config["num_labels"]
            else:
                num_labels = 3
        
        # Lấy base model name từ config
        base_model_name = None
        if hasattr(config, "_name_or_path"):
            base_model_name = config._name_or_path
        elif isinstance(config, dict) and "_name_or_path" in config:
            base_model_name = config["_name_or_path"]
        
        # Khởi tạo model với base model name
        if base_model_name:
            model = cls(model_name=base_model_name, num_labels=num_labels, **kwargs)
        else:
            # Fallback: dùng default model_name từ class
            model = cls(num_labels=num_labels, **kwargs)
        
        return model

class PhoBERTV2Model(BaseHateSpeechModel):
    """PhoBERT-V2 cho hate speech detection"""
    def __init__(self, model_name: str = "vinai/phobert-base-v2", num_labels: int = 3):
        super().__init__(model_name, num_labels)
        self.config = AutoConfig.from_pretrained(model_name, ignore_mismatched_sizes=True)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config, ignore_mismatched_sizes=True)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(self.config.hidden_size, num_labels)
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        pooled_output = outputs.pooler_output
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class BartPhoModel(BaseHateSpeechModel):
    """BART Pho cho hate speech detection"""
    def __init__(self, model_name: str = "vinai/bartpho-syllable-base", num_labels: int = 3):
        super().__init__(model_name, num_labels)
        self.config = AutoConfig.from_pretrained(model_name, ignore_mismatched_sizes=True)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config, ignore_mismatched_sizes=True)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(self.config.d_model, num_labels)
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        # Sử dụng hidden state của token cuối cùng
        last_hidden_states = outputs.last_hidden_state
        pooled_output = last_hidden_states.mean(dim=1)  # Mean pooling
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class ViSoBERTModel(BaseHateSpeechModel):
    """ViSoBERT cho hate speech detection"""
    def __init__(self, model_name: str = "uitnlp/visobert", num_labels: int = 3):
        super().__init__(model_name, num_labels)
        self.config = AutoConfig.from_pretrained(model_name, ignore_mismatched_sizes=True)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config, ignore_mismatched_sizes=True)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(self.config.hidden_size, num_labels)
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        
        # Kiểm tra xem có pooler_output không, nếu không thì dùng last_hidden_state
        if hasattr(outputs, 'pooler_output') and outputs.pooler_output is not None:
            pooled_output = outputs.pooler_output
        else:
            # Fallback: sử dụng mean pooling của last_hidden_state
            pooled_output = outputs.last_hidden_state.mean(dim=1)
            
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class PhoBERTV1Model(BaseHateSpeechModel):
    """PhoBERT-V1 cho hate speech detection"""
    def __init__(self, model_name: str = "vinai/phobert-base", num_labels: int = 3):
        super().__init__(model_name, num_labels)
        self.config = AutoConfig.from_pretrained(model_name, ignore_mismatched_sizes=True)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config, ignore_mismatched_sizes=True)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(self.config.hidden_size, num_labels)
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        # Một số encoder không có pooler_output
        if hasattr(outputs, 'pooler_output') and outputs.pooler_output is not None:
            pooled_output = outputs.pooler_output
        else:
            pooled_output = outputs.last_hidden_state.mean(dim=1)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class MBERTModel(BaseHateSpeechModel):
    """mBERT (bert-base-multilingual-cased) cho hate speech detection"""
    def __init__(self, model_name: str = "bert-base-multilingual-cased", num_labels: int = 3):
        super().__init__(model_name, num_labels)
        self.config = AutoConfig.from_pretrained(model_name, ignore_mismatched_sizes=True)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config, ignore_mismatched_sizes=True)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(self.config.hidden_size, num_labels)
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        if hasattr(outputs, 'pooler_output') and outputs.pooler_output is not None:
            pooled_output = outputs.pooler_output
        else:
            pooled_output = outputs.last_hidden_state.mean(dim=1)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class SPhoBERTModel(BaseHateSpeechModel):
    """SPhoBERT (biến thể PhoBERT syllable-level) cho hate speech detection"""
    def __init__(self, model_name: str = "vinai/phobert-base", num_labels: int = 3):
        super().__init__(model_name, num_labels)
        self.config = AutoConfig.from_pretrained(model_name, ignore_mismatched_sizes=True)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config, ignore_mismatched_sizes=True)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(self.config.hidden_size, num_labels)
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        if hasattr(outputs, 'pooler_output') and outputs.pooler_output is not None:
            pooled_output = outputs.pooler_output
        else:
            pooled_output = outputs.last_hidden_state.mean(dim=1)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class ViHateT5Model(BaseHateSpeechModel):
    """ViHateT5 cho hate speech detection"""
    def __init__(self, model_name: str = "VietAI/vit5-base", num_labels: int = 3):
        super().__init__(model_name, num_labels)
        self.config = T5Config.from_pretrained(model_name)
        self.encoder = T5ForConditionalGeneration.from_pretrained(model_name, config=self.config)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(self.config.d_model, num_labels)
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder.encoder(input_ids=input_ids, attention_mask=attention_mask)
        # Sử dụng hidden state của token cuối cùng
        last_hidden_states = outputs.last_hidden_state
        pooled_output = last_hidden_states.mean(dim=1)  # Mean pooling
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class XLMRModel(BaseHateSpeechModel):
    """XLM-R Large cho hate speech detection"""
    def __init__(self, model_name: str = "xlm-roberta-large", num_labels: int = 3):
        super().__init__(model_name, num_labels)
        self.config = AutoConfig.from_pretrained(model_name, ignore_mismatched_sizes=True)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config, ignore_mismatched_sizes=True)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(self.config.hidden_size, num_labels)
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        pooled_output = outputs.pooler_output
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class RoBERTaGRUModel(BaseHateSpeechModel):
    """RoBERTa + GRU Hybrid model"""
    def __init__(self, model_name: str = "vinai/phobert-base-v2", num_labels: int = 3, hidden_size: int = 256):
        super().__init__(model_name, num_labels)
        self.config = AutoConfig.from_pretrained(model_name, ignore_mismatched_sizes=True)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config, ignore_mismatched_sizes=True)
        self.gru = nn.GRU(
            input_size=self.config.hidden_size,
            hidden_size=hidden_size,
            num_layers=2,
            batch_first=True,
            dropout=0.1,
            bidirectional=True
        )
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(hidden_size * 2, num_labels)  # *2 for bidirectional
        
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        hidden_states = outputs.last_hidden_state  # [batch_size, seq_len, hidden_size]
        
        # GRU processing
        gru_output, _ = self.gru(hidden_states)  # [batch_size, seq_len, hidden_size*2]
        
        # Global average pooling
        pooled_output = gru_output.mean(dim=1)  # [batch_size, hidden_size*2]
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class TextCNNModel(BaseHateSpeechModel):
    """TextCNN cho hate speech detection"""
    def __init__(self, vocab_size: int, embedding_dim: int = 128, num_labels: int = 3, 
                 num_filters: int = 100, filter_sizes: list = [3, 4, 5], dropout: float = 0.5):
        super().__init__("textcnn", num_labels)
        self.embedding = nn.Embedding(vocab_size, embedding_dim)
        self.convs = nn.ModuleList([
            nn.Conv2d(1, num_filters, (filter_size, embedding_dim))
            for filter_size in filter_sizes
        ])
        self.dropout = nn.Dropout(dropout)
        self.classifier = nn.Linear(num_filters * len(filter_sizes), num_labels)
    
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
        """Override để detect vocab_size từ state_dict hoặc checkpoint file"""
        # Get vocab_size từ kwargs hoặc config
        vocab_size = kwargs.pop("vocab_size", None)
        config = kwargs.pop("config", None)
        
        # Nếu chưa có vocab_size, thử detect từ checkpoint file
        if vocab_size is None:
            import os
            state_dict = None
            # Try to load state_dict từ local path để detect vocab_size
            if os.path.isdir(pretrained_model_name_or_path):
                if os.path.isfile(os.path.join(pretrained_model_name_or_path, SAFE_WEIGHTS_NAME)):
                    try:
                        from safetensors.torch import load_file
                        state_dict = load_file(os.path.join(pretrained_model_name_or_path, SAFE_WEIGHTS_NAME))
                    except Exception:
                        pass
                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
                    try:
                        state_dict = torch.load(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME), map_location="cpu")
                    except Exception:
                        pass
            
            # Detect vocab_size từ embedding.weight
            if state_dict is not None and "embedding.weight" in state_dict:
                vocab_size = state_dict["embedding.weight"].shape[0]
            else:
                vocab_size = 30000  # Default
        
        # Get num_labels
        num_labels = kwargs.pop("num_labels", None)
        if num_labels is None:
            if config and hasattr(config, "num_labels"):
                num_labels = config.num_labels
            elif config and isinstance(config, dict) and "num_labels" in config:
                num_labels = config["num_labels"]
            else:
                num_labels = 3
        
        # Khởi tạo model
        model = cls(vocab_size=vocab_size, num_labels=num_labels, **kwargs)
        
        return model
        
    def forward(self, input_ids, attention_mask, labels=None):
        # Embedding
        embedded = self.embedding(input_ids)  # [batch_size, seq_len, embedding_dim]
        
        # Add channel dimension for Conv2d
        embedded = embedded.unsqueeze(1)  # [batch_size, 1, seq_len, embedding_dim]
        
        # Convolutional layers
        conv_outputs = []
        for conv in self.convs:
            conv_out = F.relu(conv(embedded))  # [batch_size, num_filters, seq_len', 1]
            conv_out = conv_out.squeeze(3)  # [batch_size, num_filters, seq_len']
            pooled = F.max_pool1d(conv_out, conv_out.size(2))  # [batch_size, num_filters, 1]
            pooled = pooled.squeeze(2)  # [batch_size, num_filters]
            conv_outputs.append(pooled)
        
        # Concatenate all conv outputs
        concatenated = torch.cat(conv_outputs, dim=1)  # [batch_size, num_filters * len(filter_sizes)]
        
        # Classification
        concatenated = self.dropout(concatenated)
        logits = self.classifier(concatenated)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class BiLSTMModel(BaseHateSpeechModel):
    """BiLSTM cho hate speech detection"""
    def __init__(self, vocab_size: int, embedding_dim: int = 128, hidden_size: int = 256, 
                 num_labels: int = 3, num_layers: int = 2, dropout: float = 0.5):
        super().__init__("bilstm", num_labels)
        self.embedding = nn.Embedding(vocab_size, embedding_dim)
        self.lstm = nn.LSTM(
            input_size=embedding_dim,
            hidden_size=hidden_size,
            num_layers=num_layers,
            batch_first=True,
            dropout=dropout if num_layers > 1 else 0,
            bidirectional=True
        )
        self.dropout = nn.Dropout(dropout)
        self.classifier = nn.Linear(hidden_size * 2, num_labels)  # *2 for bidirectional
    
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
        """Override để detect vocab_size từ state_dict hoặc checkpoint file"""
        # Get vocab_size từ kwargs hoặc config
        vocab_size = kwargs.pop("vocab_size", None)
        config = kwargs.pop("config", None)
        
        # Nếu chưa có vocab_size, thử detect từ checkpoint file
        if vocab_size is None:
            import os
            state_dict = None
            # Try to load state_dict từ local path để detect vocab_size
            if os.path.isdir(pretrained_model_name_or_path):
                if os.path.isfile(os.path.join(pretrained_model_name_or_path, SAFE_WEIGHTS_NAME)):
                    try:
                        from safetensors.torch import load_file
                        state_dict = load_file(os.path.join(pretrained_model_name_or_path, SAFE_WEIGHTS_NAME))
                    except Exception:
                        pass
                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
                    try:
                        state_dict = torch.load(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME), map_location="cpu")
                    except Exception:
                        pass
            
            # Detect vocab_size từ embedding.weight
            if state_dict is not None and "embedding.weight" in state_dict:
                vocab_size = state_dict["embedding.weight"].shape[0]
            else:
                vocab_size = 30000  # Default
        
        # Get num_labels
        num_labels = kwargs.pop("num_labels", None)
        if num_labels is None:
            if config and hasattr(config, "num_labels"):
                num_labels = config.num_labels
            elif config and isinstance(config, dict) and "num_labels" in config:
                num_labels = config["num_labels"]
            else:
                num_labels = 3
        
        # Khởi tạo model
        model = cls(vocab_size=vocab_size, num_labels=num_labels, **kwargs)
        
        return model
        
    def forward(self, input_ids, attention_mask, labels=None):
        # Embedding
        embedded = self.embedding(input_ids)  # [batch_size, seq_len, embedding_dim]
        
        # BiLSTM
        lstm_output, (hidden, cell) = self.lstm(embedded)  # [batch_size, seq_len, hidden_size*2]
        
        # Global average pooling (có thể thay bằng max pooling hoặc last hidden state)
        # Option 1: Global average pooling
        pooled_output = lstm_output.mean(dim=1)  # [batch_size, hidden_size*2]
        
        # Option 2: Last hidden state (uncomment if preferred)
        # pooled_output = lstm_output[:, -1, :]  # [batch_size, hidden_size*2]
        
        # Option 3: Max pooling (uncomment if preferred)
        # pooled_output = torch.max(lstm_output, dim=1)[0]  # [batch_size, hidden_size*2]
        
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss()
            loss = loss_fn(logits, labels)
        return {"loss": loss, "logits": logits}

class EnsembleModel(BaseHateSpeechModel):
    """Ensemble model kết hợp các mô hình deep learning"""
    def __init__(self, models: list, num_labels: int = 3, weights: list = None):
        super().__init__("ensemble", num_labels)
        self.models = nn.ModuleList(models)
        self.num_models = len(models)
        self.weights = weights if weights else [1.0] * self.num_models
        self.weights = torch.tensor(self.weights, dtype=torch.float32)
        self.weights = self.weights / self.weights.sum()  # Normalize weights
        
    def forward(self, input_ids, attention_mask, labels=None):
        all_logits = []
        total_loss = 0
        
        for i, model in enumerate(self.models):
            model_output = model(input_ids, attention_mask, labels)
            all_logits.append(model_output["logits"])
            
            if model_output["loss"] is not None:
                total_loss += self.weights[i] * model_output["loss"]
        
        # Weighted average of logits
        ensemble_logits = torch.zeros_like(all_logits[0])
        for i, logits in enumerate(all_logits):
            ensemble_logits += self.weights[i] * logits
            
        return {
            "loss": total_loss if total_loss > 0 else None,
            "logits": ensemble_logits
        }

def get_model(model_name: str, num_labels: int = 3, **kwargs):
    """
    Factory function để tạo model dựa trên tên
    
    Args:
        model_name: Tên model ("phobert-v2", "bartpho", "visobert", "vihate-t5", 
                   "xlm-r", "roberta-gru", "textcnn", "bilstm", "bilstm-crf", "ensemble")
        num_labels: Số lượng nhãn (3 cho hate speech: CLEAN, OFFENSIVE, HATE)
        **kwargs: Các tham số bổ sung cho model
    
    Returns:
        Model instance
    """
    model_mapping = {
        "phobert-v1": PhoBERTV1Model,
        "phobert-v2": PhoBERTV2Model,
        "bartpho": BartPhoModel,
        "visobert": ViSoBERTModel,
        "vihate-t5": ViHateT5Model,
        "xlm-r": XLMRModel,
        "mbert": MBERTModel,
        "sphobert": SPhoBERTModel,
        "roberta-gru": RoBERTaGRUModel,
        "textcnn": TextCNNModel,
        "bilstm": BiLSTMModel,
        "ensemble": EnsembleModel
    }
    
    if model_name not in model_mapping:
        raise ValueError(f"Unknown model: {model_name}. Available models: {list(model_mapping.keys())}")
    
    model_class = model_mapping[model_name]
    return model_class(num_labels=num_labels, **kwargs)