deployment/detection_trt/convert_model.py

#!/usr/bin/env python3
"""
模型转换脚本 - PyTorch -> ONNX -> TensorRT
支持 DeCLIP (csa模式) 和 CLIP (vanilla模式) 的转换

Usage:
    python convert_model.py --model-type declip --mode csa --checkpoint <path>
    python convert_model.py --model-type clip --mode vanilla --checkpoint <path>
"""

import os
import sys
import argparse
import time
from pathlib import Path

import torch
import torch.onnx
import numpy as np

# 添加项目根目录到路径
SCRIPT_DIR = Path(__file__).parent
DECLIP_ROOT = SCRIPT_DIR.parent
sys.path.insert(0, str(DECLIP_ROOT))
sys.path.insert(0, str(DECLIP_ROOT / 'src'))

from configs.custom_modules import (
    prepare_model_for_export,
    register_custom_rewriters,
    disable_xformers_for_export
)


def parse_args():
    parser = argparse.ArgumentParser(description='DeCLIP/CLIP TensorRT 转换')
    
    # 模型配置
    parser.add_argument('--model-type', type=str, default='declip',
                        choices=['declip', 'clip'],
                        help='模型类型: declip (DeCLIP) 或 clip (原始CLIP)')
    parser.add_argument('--model-name', type=str, default='EVA02-CLIP-B-16',
                        help='模型名称')
    parser.add_argument('--checkpoint', type=str, required=True,
                        help='模型检查点路径')
    parser.add_argument('--mode', type=str, default='csa',
                        choices=['vanilla', 'csa', 'qq', 'kk'],
                        help='特征模式: vanilla (原始) 或 csa (DeCLIP)')
    
    # 输入配置
    parser.add_argument('--image-size', type=int, default=560,
                        help='输入图像尺寸')
    parser.add_argument('--batch-size', type=int, default=1,
                        help='批处理大小')
    
    # 输出配置
    parser.add_argument('--output-dir', type=str, default='engines',
                        help='输出目录')
    parser.add_argument('--output-name', type=str, default=None,
                        help='输出文件名 (不含扩展名)')
    
    # TRT 配置
    parser.add_argument('--fp16', action='store_true', default=True,
                        help='使用 FP16 精度')
    parser.add_argument('--int8', action='store_true',
                        help='使用 INT8 精度 (需要校准数据)')
    parser.add_argument('--workspace', type=int, default=4,
                        help='TRT workspace 大小 (GB)')
    
    # 动态形状
    parser.add_argument('--dynamic', action='store_true',
                        help='启用动态输入形状')
    parser.add_argument('--min-shape', type=int, nargs=2, default=[560, 560],
                        help='最小输入尺寸 [H, W]')
    parser.add_argument('--max-shape', type=int, nargs=2, default=[800, 1333],
                        help='最大输入尺寸 [H, W]')
    
    # 其他
    parser.add_argument('--device', type=str, default='cuda:0',
                        help='设备')
    parser.add_argument('--verify', action='store_true', default=True,
                        help='验证转换后的模型精度')
    parser.add_argument('--verbose', action='store_true',
                        help='详细输出')
    
    return parser.parse_args()


def load_model(args):
    """加载 EVA-CLIP 模型"""
    print(f"Loading model: {args.model_name}")
    print(f"Checkpoint: {args.checkpoint}")
    
    from open_clip import create_model
    
    # 创建模型
    model = create_model(
        args.model_name,
        pretrained='eva',
        device=args.device,
        precision='fp32',
        output_dict=True,
        cache_dir=args.checkpoint
    )
    
    # 加载检查点 (如果是微调后的权重)
    if os.path.isfile(args.checkpoint):
        print(f"Loading checkpoint from {args.checkpoint}")
        checkpoint = torch.load(args.checkpoint, map_location=args.device)
        if 'state_dict' in checkpoint:
            model.load_state_dict(checkpoint['state_dict'], strict=False)
        elif 'model' in checkpoint:
            model.load_state_dict(checkpoint['model'], strict=False)
        else:
            model.load_state_dict(checkpoint, strict=False)
    
    model.eval()
    return model


def export_to_onnx(model, args):
    """导出模型到 ONNX 格式"""
    output_dir = Path(args.output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    
    # 生成输出文件名
    if args.output_name:
        output_name = args.output_name
    else:
        output_name = f"{args.model_type}_{args.mode}_{args.image_size}"
    
    onnx_path = output_dir / f"{output_name}.onnx"
    
    print(f"\n{'='*60}")
    print(f"Exporting to ONNX: {onnx_path}")
    print(f"{'='*60}")
    
    # 准备模型
    model = prepare_model_for_export(model, mode=args.mode)
    model = model.to(args.device)
    
    # 创建示例输入
    dummy_input = torch.randn(
        args.batch_size, 3, args.image_size, args.image_size,
        device=args.device,
        dtype=torch.float32
    )
    
    # 动态轴配置
    if args.dynamic:
        dynamic_axes = {
            'input': {0: 'batch', 2: 'height', 3: 'width'},
            'output': {0: 'batch', 2: 'feat_height', 3: 'feat_width'}
        }
    else:
        dynamic_axes = None
    
    # 导出 ONNX
    torch.onnx.export(
        model,
        dummy_input,
        str(onnx_path),
        input_names=['input'],
        output_names=['output'],
        dynamic_axes=dynamic_axes,
        opset_version=17,
        do_constant_folding=True,
        verbose=args.verbose
    )
    
    print(f"ONNX model saved to: {onnx_path}")
    
    # 简化 ONNX 模型
    try:
        import onnx
        from onnxsim import simplify
        
        print("Simplifying ONNX model...")
        onnx_model = onnx.load(str(onnx_path))
        simplified_model, check = simplify(onnx_model)
        
        if check:
            simplified_path = output_dir / f"{output_name}_simplified.onnx"
            onnx.save(simplified_model, str(simplified_path))
            print(f"Simplified ONNX model saved to: {simplified_path}")
            return str(simplified_path)
        else:
            print("Warning: ONNX simplification failed, using original model")
    except ImportError:
        print("onnxsim not installed, skipping simplification")
    
    return str(onnx_path)


def convert_to_tensorrt(onnx_path, args):
    """转换 ONNX 到 TensorRT 引擎"""
    try:
        import tensorrt as trt
    except ImportError:
        print("Error: TensorRT not installed")
        return None
    
    output_dir = Path(args.output_dir)
    engine_name = Path(onnx_path).stem.replace('_simplified', '')
    
    if args.fp16:
        engine_name += '_fp16'
    if args.int8:
        engine_name += '_int8'
    
    engine_path = output_dir / f"{engine_name}.engine"
    
    print(f"\n{'='*60}")
    print(f"Converting to TensorRT: {engine_path}")
    print(f"{'='*60}")
    
    # 创建 TRT builder
    logger = trt.Logger(trt.Logger.VERBOSE if args.verbose else trt.Logger.WARNING)
    builder = trt.Builder(logger)
    network = builder.create_network(
        1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
    )
    parser = trt.OnnxParser(network, logger)
    
    # 解析 ONNX
    print("Parsing ONNX model...")
    with open(onnx_path, 'rb') as f:
        if not parser.parse(f.read()):
            for error in range(parser.num_errors):
                print(f"ONNX Parse Error: {parser.get_error(error)}")
            return None
    
    # 创建 builder config
    config = builder.create_builder_config()
    config.set_memory_pool_limit(
        trt.MemoryPoolType.WORKSPACE,
        args.workspace * (1 << 30)  # GB to bytes
    )
    
    # 精度设置
    if args.fp16:
        if builder.platform_has_fast_fp16:
            config.set_flag(trt.BuilderFlag.FP16)
            print("FP16 mode enabled")
        else:
            print("Warning: FP16 not supported on this platform")
    
    if args.int8:
        if builder.platform_has_fast_int8:
            config.set_flag(trt.BuilderFlag.INT8)
            print("INT8 mode enabled (requires calibration)")
        else:
            print("Warning: INT8 not supported on this platform")
    
    # 动态形状配置
    if args.dynamic:
        profile = builder.create_optimization_profile()
        input_tensor = network.get_input(0)
        
        min_shape = (args.batch_size, 3, args.min_shape[0], args.min_shape[1])
        opt_shape = (args.batch_size, 3, args.image_size, args.image_size)
        max_shape = (args.batch_size, 3, args.max_shape[0], args.max_shape[1])
        
        profile.set_shape(input_tensor.name, min_shape, opt_shape, max_shape)
        config.add_optimization_profile(profile)
        print(f"Dynamic shapes: min={min_shape}, opt={opt_shape}, max={max_shape}")
    
    # 构建引擎
    print("Building TensorRT engine (this may take a while)...")
    start_time = time.time()
    
    serialized_engine = builder.build_serialized_network(network, config)
    
    if serialized_engine is None:
        print("Error: Failed to build TensorRT engine")
        return None
    
    build_time = time.time() - start_time
    print(f"Engine built in {build_time:.1f} seconds")
    
    # 保存引擎
    with open(engine_path, 'wb') as f:
        f.write(serialized_engine)
    
    print(f"TensorRT engine saved to: {engine_path}")
    
    # 打印引擎信息
    runtime = trt.Runtime(logger)
    engine = runtime.deserialize_cuda_engine(serialized_engine)
    
    print(f"\nEngine Info:")
    print(f"  - Number of layers: {engine.num_layers}")
    print(f"  - Number of IO tensors: {engine.num_io_tensors}")
    
    for i in range(engine.num_io_tensors):
        name = engine.get_tensor_name(i)
        shape = engine.get_tensor_shape(name)
        dtype = engine.get_tensor_dtype(name)
        mode = engine.get_tensor_mode(name)
        print(f"  - {name}: shape={shape}, dtype={dtype}, mode={mode}")
    
    return str(engine_path)


def verify_model(pytorch_model, onnx_path, engine_path, args):
    """验证转换后的模型精度"""
    print(f"\n{'='*60}")
    print("Verifying model accuracy")
    print(f"{'='*60}")
    
    # 准备输入
    dummy_input = torch.randn(
        1, 3, args.image_size, args.image_size,
        device=args.device,
        dtype=torch.float32
    )
    
    # PyTorch 推理
    pytorch_model = prepare_model_for_export(pytorch_model, mode=args.mode)
    pytorch_model = pytorch_model.to(args.device).eval()
    
    with torch.no_grad():
        pytorch_output = pytorch_model(dummy_input)
    
    pytorch_output = pytorch_output.cpu().numpy()
    print(f"PyTorch output shape: {pytorch_output.shape}")
    
    # ONNX 推理
    try:
        import onnxruntime as ort
        
        ort_session = ort.InferenceSession(
            onnx_path,
            providers=['CUDAExecutionProvider', 'CPUExecutionProvider']
        )
        
        onnx_output = ort_session.run(
            None,
            {'input': dummy_input.cpu().numpy()}
        )[0]
        
        # 计算误差
        onnx_diff = np.abs(pytorch_output - onnx_output).max()
        print(f"ONNX max difference: {onnx_diff:.6f}")
        
        if onnx_diff < 1e-3:
            print("✓ ONNX verification passed")
        else:
            print("✗ ONNX verification failed (diff > 1e-3)")
            
    except ImportError:
        print("ONNX Runtime not installed, skipping ONNX verification")
    
    # TensorRT 推理
    if engine_path and os.path.exists(engine_path):
        try:
            import tensorrt as trt
            import pycuda.driver as cuda
            import pycuda.autoinit
            
            # 加载引擎
            logger = trt.Logger(trt.Logger.WARNING)
            runtime = trt.Runtime(logger)
            
            with open(engine_path, 'rb') as f:
                engine = runtime.deserialize_cuda_engine(f.read())
            
            context = engine.create_execution_context()
            
            # 分配内存
            input_name = engine.get_tensor_name(0)
            output_name = engine.get_tensor_name(1)
            
            input_shape = list(dummy_input.shape)
            context.set_input_shape(input_name, input_shape)
            
            output_shape = context.get_tensor_shape(output_name)
            output_size = int(np.prod(output_shape))
            
            # GPU 内存
            d_input = cuda.mem_alloc(dummy_input.numpy().nbytes)
            d_output = cuda.mem_alloc(output_size * np.float32().nbytes)
            
            # 复制输入
            cuda.memcpy_htod(d_input, dummy_input.cpu().numpy())
            
            # 设置绑定
            context.set_tensor_address(input_name, int(d_input))
            context.set_tensor_address(output_name, int(d_output))
            
            # 推理
            stream = cuda.Stream()
            context.execute_async_v3(stream.handle)
            stream.synchronize()
            
            # 复制输出
            trt_output = np.empty(output_shape, dtype=np.float32)
            cuda.memcpy_dtoh(trt_output, d_output)
            
            # 计算误差
            trt_diff = np.abs(pytorch_output - trt_output).max()
            print(f"TensorRT max difference: {trt_diff:.6f}")
            
            if trt_diff < 1e-2:  # TRT FP16 允许更大误差
                print("✓ TensorRT verification passed")
            else:
                print("✗ TensorRT verification failed (diff > 1e-2)")
                
        except ImportError as e:
            print(f"TensorRT/PyCUDA not installed, skipping TRT verification: {e}")
        except Exception as e:
            print(f"TensorRT verification error: {e}")


def main():
    args = parse_args()
    
    print(f"\n{'='*60}")
    print("DeCLIP/CLIP TensorRT Conversion")
    print(f"{'='*60}")
    print(f"Model Type: {args.model_type}")
    print(f"Mode: {args.mode}")
    print(f"Image Size: {args.image_size}")
    print(f"FP16: {args.fp16}")
    print(f"Dynamic Shapes: {args.dynamic}")
    print(f"{'='*60}\n")
    
    # 注册自定义重写器
    register_custom_rewriters()
    
    # 加载模型
    model = load_model(args)
    
    # 导出 ONNX
    onnx_path = export_to_onnx(model, args)
    
    # 转换 TensorRT
    engine_path = convert_to_tensorrt(onnx_path, args)
    
    # 验证精度
    if args.verify and engine_path:
        verify_model(model, onnx_path, engine_path, args)
    
    print(f"\n{'='*60}")
    print("Conversion complete!")
    print(f"{'='*60}")
    print(f"ONNX: {onnx_path}")
    if engine_path:
        print(f"TensorRT: {engine_path}")
    print()


if __name__ == '__main__':
    main()