pipeline.py

from transformers import PreTrainedModel, PretrainedConfig
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing.text import tokenizer_from_json
from tensorflow.keras.preprocessing.sequence import pad_sequences
import numpy as np
import json

class NewsClassifierConfig(PretrainedConfig):
    model_type = "custom"
    
    def __init__(
        self,
        max_length=41,  # Modified to match model input shape
        vocab_size=74934,  # Modified based on embedding layer size
        embedding_dim=128,  # Added to match model architecture
        hidden_size=64,  # Matches final LSTM layer
        num_labels=2,
        **kwargs
    ):
        self.max_length = max_length
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.hidden_size = hidden_size
        self.num_labels = num_labels
        super().__init__(**kwargs)

class NewsClassifier(PreTrainedModel):
    config_class = NewsClassifierConfig
    base_model_prefix = "custom"

    def __init__(self, config):
        super().__init__(config)
        self.model = None  # Will be loaded in post_init
        self.tokenizer = None
        
    def post_init(self):
        """Load model and tokenizer after initialization"""
        self.model = load_model('custom.h5')
        with open('tokenizer.json', 'r') as f:
            tokenizer_data = json.load(f)
            self.tokenizer = tokenizer_from_json(tokenizer_data)

    def forward(self, text_input):
        if not self.model or not self.tokenizer:
            self.post_init()
            
        if isinstance(text_input, str):
            text_input = [text_input]
            
        sequences = self.tokenizer.texts_to_sequences(text_input)
        padded = pad_sequences(sequences, maxlen=self.config.max_length)
        predictions = self.model.predict(padded, verbose=0)
        
        results = []
        for pred in predictions:
            # Convert from 2-class output to single score
            score = float(pred[1])  # Assuming pred[1] is the probability for "foxnews"
            label = "foxnews" if score > 0.5 else "nbc"
            results.append({
                "label": label,
                "score": score if label == "foxnews" else 1 - score
            })
        
        return results[0] if len(text_input) == 1 else results