convert_onnx_batch_size.py

#!/usr/bin/env python3
"""
将已导出的ONNX模型转换为支持指定batch_size的模型
"""

import onnx
import onnx.shape_inference
import argparse
import os

def modify_onnx_batch_size(model_path, output_path, target_batch_size=24):
    """修改ONNX模型的batch_size
    
    Args:
        model_path: 输入模型路径
        output_path: 输出模型路径
        target_batch_size: 目标batch_size，-1表示动态batch，其他值表示固定batch
    """
    
    # 加载模型
    model = onnx.load(model_path)
    
    # 获取模型输入信息
    print(f"原始模型输入信息:")
    for i, input_info in enumerate(model.graph.input):
        print(f"  Input {i}: {input_info.name}")
        if input_info.type.tensor_type.HasField("shape"):
            shape = input_info.type.tensor_type.shape
            print(f"    原始形状: ", end="")
            for j, dim in enumerate(shape.dim):
                if dim.HasField("dim_value"):
                    print(f"{dim.dim_value}", end=" ")
                elif dim.HasField("dim_param"):
                    print(f"{dim.dim_param}", end=" ")
            print()
    
    # 修改输入形状
    for input_info in model.graph.input:
        if input_info.type.tensor_type.HasField("shape"):
            shape = input_info.type.tensor_type.shape
            # 修改第一个维度（batch_size）
            if len(shape.dim) > 0:
                if target_batch_size == -1:
                    # 动态batch_size模式
                    if shape.dim[0].HasField("dim_value"):
                        shape.dim[0].dim_param = "batch_size"
                        shape.dim[0].ClearField("dim_value")
                    elif shape.dim[0].HasField("dim_param"):
                        # 已经是动态维度，保持不变
                        pass
                    else:
                        # 其他情况，设为动态维度
                        shape.dim[0].dim_param = "batch_size"
                else:
                    # 固定batch_size模式
                    shape.dim[0].dim_value = target_batch_size
                    if shape.dim[0].HasField("dim_param"):
                        shape.dim[0].ClearField("dim_param")
    
    # 修改输出形状
    for output_info in model.graph.output:
        if output_info.type.tensor_type.HasField("shape"):
            shape = output_info.type.tensor_type.shape
            if len(shape.dim) > 0:
                if target_batch_size == -1:
                    # 动态batch_size模式
                    if shape.dim[0].HasField("dim_value"):
                        shape.dim[0].dim_param = "batch_size"
                        shape.dim[0].ClearField("dim_value")
                else:
                    # 固定batch_size模式
                    if shape.dim[0].HasField("dim_value"):
                        shape.dim[0].dim_value = target_batch_size
                    elif shape.dim[0].HasField("dim_param"):
                        shape.dim[0].ClearField("dim_param")
                        shape.dim[0].dim_value = target_batch_size
    
    # 运行形状推断
    model = onnx.shape_inference.infer_shapes(model)
    
    # 保存修改后的模型
    onnx.save(model, output_path)
    
    print(f"模型已保存到: {output_path}")
    print(f"目标batch_size: {'动态' if target_batch_size == -1 else target_batch_size}")
    
    # 验证修改结果
    print(f"修改后模型输入信息:")
    model_modified = onnx.load(output_path)
    for i, input_info in enumerate(model_modified.graph.input):
        print(f"  Input {i}: {input_info.name}")
        if input_info.type.tensor_type.HasField("shape"):
            shape = input_info.type.tensor_type.shape
            print(f"    修改后形状: ", end="")
            for j, dim in enumerate(shape.dim):
                if dim.HasField("dim_value"):
                    print(f"{dim.dim_value}", end=" ")
                elif dim.HasField("dim_param"):
                    print(f"{dim.dim_param}", end=" ")
            print()

def batch_convert_models(input_dir, output_dir, target_batch_size=24):
    """批量转换目录中的所有ONNX模型"""
    
    if not os.path.exists(output_dir):
        os.makedirs(output_dir)
    
    onnx_files = [f for f in os.listdir(input_dir) if f.endswith('.onnx')]
    
    print(f"找到 {len(onnx_files)} 个ONNX文件:")
    for file in onnx_files:
        print(f"  - {file}")
    
    for file in onnx_files:
        input_path = os.path.join(input_dir, file)
        output_path = os.path.join(output_dir, file)
        
        print(f"\n正在转换: {file}")
        try:
            modify_onnx_batch_size(input_path, output_path, target_batch_size)
            print(f"✓ {file} 转换成功")
        except Exception as e:
            print(f"✗ {file} 转换失败: {e}")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='修改ONNX模型的batch_size')
    parser.add_argument('--input', type=str, required=True, help='输入ONNX文件或目录路径')
    parser.add_argument('--output', type=str, required=True, help='输出路径')
    parser.add_argument('--batch_size', type=int, default=24, help='目标batch_size（-1表示动态batch）')
    parser.add_argument('--batch_mode', action='store_true', help='批量模式，处理目录中的所有ONNX文件')
    
    args = parser.parse_args()
    
    if args.batch_mode:
        # 批量模式
        batch_convert_models(args.input, args.output, args.batch_size)
    else:
        # 单个文件模式
        modify_onnx_batch_size(args.input, args.output, args.batch_size)