#!/usr/bin/env python3 """ Export trained CodeBERT model to ONNX format with optional quantization. Supports both CPU and GPU inference. """ import os import sys import torch import torch.nn as nn from transformers import RobertaTokenizer, RobertaModel import json # Paths MODEL_PATH = "/c1/new-models/best_model.pt" CODEBERT_PATH = "/c1/huggingface/codebert-base" OUTPUT_DIR = "/c1/new-models" ONNX_PATH = os.path.join(OUTPUT_DIR, "model.onnx") ONNX_QUANTIZED_PATH = os.path.join(OUTPUT_DIR, "model_quantized.onnx") class CodeBERTClassifier(nn.Module): """CodeBERT-based classifier for web attack detection - matches training script.""" def __init__(self, model_path, num_labels=2, dropout=0.1): super(CodeBERTClassifier, self).__init__() self.codebert = RobertaModel.from_pretrained(model_path) self.dropout = nn.Dropout(dropout) self.classifier = nn.Linear(self.codebert.config.hidden_size, num_labels) def forward(self, input_ids, attention_mask): outputs = self.codebert(input_ids=input_ids, attention_mask=attention_mask) pooled_output = outputs.pooler_output pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_output) return logits class ONNXCodeBERTClassifier(nn.Module): """Wrapper for ONNX export with softmax output.""" def __init__(self, model): super().__init__() self.model = model self.model.dropout.p = 0 # Disable dropout for inference def forward(self, input_ids, attention_mask): outputs = self.model.codebert(input_ids=input_ids, attention_mask=attention_mask) pooled_output = outputs.pooler_output logits = self.model.classifier(pooled_output) probabilities = torch.softmax(logits, dim=-1) return probabilities def export_to_onnx(): """Export model to ONNX format.""" print("=" * 80) print("ONNX Model Export") print("=" * 80) # Device - use CPU for export to avoid CUDA issues device = torch.device("cpu") print(f"Export Device: {device}") # Load tokenizer print("\n1. Loading tokenizer...") tokenizer = RobertaTokenizer.from_pretrained(CODEBERT_PATH) print(f" Tokenizer loaded: {type(tokenizer).__name__}") # Create model with same architecture as training print("\n2. Loading model...") model = CodeBERTClassifier(CODEBERT_PATH) # Load trained weights checkpoint = torch.load(MODEL_PATH, map_location=device) if isinstance(checkpoint, dict) and 'model_state_dict' in checkpoint: model.load_state_dict(checkpoint['model_state_dict']) else: model.load_state_dict(checkpoint) model.eval() model.to(device) print(f" Model loaded from: {MODEL_PATH}") # Wrap for ONNX export onnx_model = ONNXCodeBERTClassifier(model) onnx_model.eval() onnx_model.to(device) # Create dummy input print("\n3. Creating dummy input...") max_length = 256 dummy_text = "SELECT * FROM users WHERE id=1" inputs = tokenizer( dummy_text, max_length=max_length, padding='max_length', truncation=True, return_tensors='pt' ) dummy_input_ids = inputs['input_ids'].to(device) dummy_attention_mask = inputs['attention_mask'].to(device) print(f" Input shape: {dummy_input_ids.shape}") # Test forward pass first print("\n4. Testing forward pass...") with torch.no_grad(): test_output = onnx_model(dummy_input_ids, dummy_attention_mask) print(f" Output shape: {test_output.shape}") print(f" Output sample: {test_output[0].numpy()}") # Export to ONNX print("\n5. Exporting to ONNX...") torch.onnx.export( onnx_model, (dummy_input_ids, dummy_attention_mask), ONNX_PATH, export_params=True, opset_version=14, do_constant_folding=True, input_names=['input_ids', 'attention_mask'], output_names=['probabilities'], dynamic_axes={ 'input_ids': {0: 'batch_size', 1: 'sequence_length'}, 'attention_mask': {0: 'batch_size', 1: 'sequence_length'}, 'probabilities': {0: 'batch_size'} } ) onnx_size = os.path.getsize(ONNX_PATH) / (1024 * 1024) print(f" ONNX model saved: {ONNX_PATH}") print(f" Size: {onnx_size:.2f} MB") # Quantize model print("\n6. Quantizing model (dynamic quantization)...") try: from onnxruntime.quantization import quantize_dynamic, QuantType quantize_dynamic( model_input=ONNX_PATH, model_output=ONNX_QUANTIZED_PATH, weight_type=QuantType.QUInt8, optimize_model=True ) quantized_size = os.path.getsize(ONNX_QUANTIZED_PATH) / (1024 * 1024) print(f" Quantized model saved: {ONNX_QUANTIZED_PATH}") print(f" Size: {quantized_size:.2f} MB") print(f" Compression ratio: {onnx_size / quantized_size:.2f}x") except Exception as e: print(f" Warning: Quantization failed: {e}") print(" Using non-quantized model.") import shutil shutil.copy(ONNX_PATH, ONNX_QUANTIZED_PATH) # Verify ONNX model print("\n7. Verifying ONNX model...") try: import onnx onnx_check = onnx.load(ONNX_PATH) onnx.checker.check_model(onnx_check) print(" ONNX model verification: PASSED") except Exception as e: print(f" Warning: ONNX verification failed: {e}") # Test inference with ONNX Runtime print("\n8. Testing ONNX Runtime inference...") try: import onnxruntime as ort import numpy as np # Try GPU first, fallback to CPU providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] available_providers = ort.get_available_providers() use_providers = [p for p in providers if p in available_providers] session = ort.InferenceSession(ONNX_PATH, providers=use_providers) actual_provider = session.get_providers()[0] print(f" Using provider: {actual_provider}") # Test inference test_texts = [ "SELECT * FROM users WHERE id=1 OR 1=1", # SQL injection "GET /index.html HTTP/1.1", # Normal request "", # XSS "Mozilla/5.0 (Windows NT 10.0; Win64)", # Normal UA ] print("\n Test predictions:") for text in test_texts: inputs = tokenizer( text, max_length=max_length, padding='max_length', truncation=True, return_tensors='np' ) outputs = session.run( None, { 'input_ids': inputs['input_ids'].astype(np.int64), 'attention_mask': inputs['attention_mask'].astype(np.int64) } ) probs = outputs[0][0] pred = np.argmax(probs) label = "Malicious" if pred == 1 else "Benign" conf = probs[pred] * 100 print(f" - '{text[:40]:<40}' => {label:<10} ({conf:.1f}%)") except Exception as e: print(f" Warning: ONNX Runtime test failed: {e}") import traceback traceback.print_exc() # Save export config print("\n9. Saving export configuration...") export_config = { "model_path": ONNX_PATH, "quantized_model_path": ONNX_QUANTIZED_PATH, "max_length": max_length, "tokenizer_path": CODEBERT_PATH, "labels": {"0": "benign", "1": "malicious"}, "input_names": ["input_ids", "attention_mask"], "output_names": ["probabilities"] } config_path = os.path.join(OUTPUT_DIR, "onnx_config.json") with open(config_path, 'w') as f: json.dump(export_config, f, indent=2) print(f" Config saved: {config_path}") print("\n" + "=" * 80) print("Export completed!") print("=" * 80) print(f"ONNX Model: {ONNX_PATH} ({onnx_size:.2f} MB)") if os.path.exists(ONNX_QUANTIZED_PATH): qsize = os.path.getsize(ONNX_QUANTIZED_PATH) / (1024 * 1024) print(f"Quantized Model: {ONNX_QUANTIZED_PATH} ({qsize:.2f} MB)") print("=" * 80) if __name__ == "__main__": export_to_onnx()