Upload inference.py with huggingface_hub

inference.py

@@ -0,0 +1,127 @@
+"""
+Inference script for multimodal fraudulent paper detection.
+"""
+
+import os
+import sys
+import torch
+import numpy as np
+from transformers import AutoTokenizer
+import argparse
+import json
+
+from model import MultimodalFraudDetector
+
+
+def predict_fraud(model, tokenizer, text, tabular, metadata, device):
+    """Predict fraud probability for a single paper."""
+    model.eval()
+    
+    # Tokenize text
+    encoding = tokenizer(
+        text,
+        max_length=512,
+        padding='max_length',
+        truncation=True,
+        return_tensors='pt'
+    )
+    
+    input_ids = encoding['input_ids'].to(device)
+    attention_mask = encoding['attention_mask'].to(device)
+    tabular = torch.tensor(tabular, dtype=torch.float32).unsqueeze(0).to(device)
+    metadata = torch.tensor(metadata, dtype=torch.float32).unsqueeze(0).to(device)
+    
+    with torch.no_grad():
+        outputs = model(
+            text_input_ids=input_ids,
+            text_attention_mask=attention_mask,
+            tabular_features=tabular,
+            metadata_features=metadata
+        )
+    
+    logits = outputs['logits']
+    probs = torch.softmax(logits, dim=1)
+    fraud_prob = probs[0, 1].item()
+    
+    modality_scores = outputs['modality_scores'][0].cpu().numpy()
+    anomaly_score = outputs['anomaly_score'][0].item()
+    
+    return {
+        'fraud_probability': fraud_prob,
+        'is_fraudulent': fraud_prob > 0.5,
+        'modality_contributions': {
+            'text': float(modality_scores[0]),
+            'image': float(modality_scores[1]),
+            'tabular': float(modality_scores[2]),
+            'metadata': float(modality_scores[3])
+        },
+        'anomaly_score': anomaly_score
+    }
+
+
+def explain_prediction(result):
+    """Generate human-readable explanation."""
+    explanations = []
+    
+    if result['fraud_probability'] > 0.5:
+        explanations.append(f"FRAUDULENT (probability: {result['fraud_probability']:.2%})")
+    else:
+        explanations.append(f"AUTHENTIC (fraud probability: {result['fraud_probability']:.2%})")
+    
+    # Modality contributions
+    contrib = result['modality_contributions']
+    max_modality = max(contrib, key=contrib.get)
+    explanations.append(f"Primary fraud indicator: {max_modality} modality (score: {contrib[max_modality]:.3f})")
+    
+    if result['anomaly_score'] > 0.7:
+        explanations.append(f"High anomaly score ({result['anomaly_score']:.3f}): Paper shows strong outlier patterns")
+    
+    return "\n".join(explanations)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model_path', required=True)
+    parser.add_argument('--text', default='')
+    parser.add_argument('--title', default='')
+    parser.add_argument('--output', default='prediction.json')
+    args = parser.parse_args()
+    
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    
+    # Load model
+    checkpoint = torch.load(args.model_path, map_location=device)
+    model_args = checkpoint.get('args', {})
+    
+    model = MultimodalFraudDetector(
+        text_model=model_args.get('text_model', 'allenai/scibert_scivocab_uncased'),
+        tabular_features=10,
+        metadata_features=12
+    ).to(device)
+    
+    model.load_state_dict(checkpoint['model_state_dict'])
+    model.eval()
+    
+    tokenizer = AutoTokenizer.from_pretrained(model_args.get('text_model'))
+    
+    # Prepare input
+    text = f"{args.title} [SEP] {args.text}"
+    
+    # Dummy features for demo (in production, extract from actual paper)
+    tabular = np.random.randn(10).astype(np.float32)
+    metadata = np.random.randn(12).astype(np.float32)
+    
+    # Predict
+    result = predict_fraud(model, tokenizer, text, tabular, metadata, device)
+    result['explanation'] = explain_prediction(result)
+    
+    print(result['explanation'])
+    
+    with open(args.output, 'w') as f:
+        json.dump(result, f, indent=2)
+    
+    print(f"\nSaved to {args.output}")
+
+
+if __name__ == '__main__':
+    main()