predict_ft_model_crosssectional.py

def predict_with_threshold(text, model_dir="cross-sectional-classifier", threshold=None, threshold_type=None):
    """
    Make a prediction using the saved model and custom threshold.
    
    Args:
        text: Input text to classify
        model_dir: Directory where the model and threshold are saved
        threshold: Custom threshold to use (if None, loads from saved config)
        threshold_type: Type of threshold to use ('f1' or 'balanced') if loading from all_thresholds.json
        
    Returns:
        Dictionary with prediction results
    """
    # Load the model and tokenizer
    from transformers import AutoModelForSequenceClassification, AutoTokenizer
    import torch
    import json
    
    tokenizer = AutoTokenizer.from_pretrained(model_dir)
    model = AutoModelForSequenceClassification.from_pretrained(model_dir)
    
    threshold_source = "custom"
    
    # Load the threshold if not provided
    if threshold is None:
        # Try multiple methods to get the threshold
        if threshold_type is not None:
            # First try all_thresholds.json if threshold_type is specified
            try:
                with open(f"{model_dir}/all_thresholds.json", "r") as f:
                    all_thresholds = json.load(f)
                    if threshold_type in all_thresholds:
                        threshold = all_thresholds[threshold_type]["threshold"]
                        threshold_source = f"all_thresholds.json ({threshold_type})"
                    else:
                        print(f"Threshold type '{threshold_type}' not found. Available types: {list(all_thresholds.keys())}")
            except FileNotFoundError:
                pass
        
        # If still no threshold, try threshold_config.json
        if threshold is None:
            try:
                with open(f"{model_dir}/threshold_config.json", "r") as f:
                    config = json.load(f)
                    threshold = config["threshold"]
                    threshold_source = "threshold_config.json"
            except FileNotFoundError:
                # Default to 0.5 if no threshold config is found
                threshold = 0.5
                threshold_source = "default"
                print("No threshold configuration found. Using default threshold of 0.5.")
    
    # Tokenize the input text
    inputs = tokenizer(text, padding="max_length", truncation=True, max_length=512, return_tensors="pt")
    
    # Get model prediction
    model.eval()
    with torch.no_grad():
        outputs = model(**inputs)
    
    # Convert logits to probabilities
    probs = torch.nn.functional.softmax(outputs.logits, dim=1).squeeze().tolist()
    
    # Apply threshold to get the final prediction
    if isinstance(probs, list):  # Handle batch size of 1
        prediction = 1 if probs[1] > threshold else 0
        class_probs = {
            "Cross-sectional": probs[1],
            "Other": probs[0]
        }
    else:  # Handle single prediction case
        prediction = 1 if probs > threshold else 0
        class_probs = {
            "Cross-sectional": probs,
            "Other": 1 - probs
        }
    
    # Map prediction back to original label
    label_names = {1: "Cross-sectional", 0: "Other"}
    
    return {
        "prediction": label_names[prediction],
        "probabilities": class_probs,
        "threshold_used": threshold,
        "threshold_source": threshold_source
    }


sample_text = 'This is a cross-sectional study that aims to investigate the relationship between smoking and lung cancer.'
result = predict_with_threshold(sample_text)