model_watermark_detect/tests/prune_tool.py

"""
对onnx权重文件进行规则剪枝
"""
import argparse
import os

import onnx
import numpy as np
from onnx import numpy_helper

def prune_weights(model_path, pruned_model, pruning_percentage=0.05):
    model = onnx.load(model_path)
    # 获取所有权重的初始化器
    weight_initializers = [init for init in model.graph.initializer if
                           init.name in {node.input[1] for node in model.graph.node if node.op_type == 'Conv'}]

    # 收集所有权重的绝对值
    all_weights = np.concatenate([numpy_helper.to_array(init).flatten() for init in weight_initializers])

    # 计算阈值
    num_weights_to_prune = int(len(all_weights) * pruning_percentage)
    threshold = np.partition(np.abs(all_weights), num_weights_to_prune)[num_weights_to_prune]

    # 剪枝权重并更新初始化器
    new_initializers = []
    for init in model.graph.initializer:
        if init.name in {node.input[1] for node in model.graph.node if node.op_type == 'Conv'}:
            weights = numpy_helper.to_array(init).copy()
            # 根据阈值剪枝权重
            weights[np.abs(weights) < threshold] = 0
            new_initializer = numpy_helper.from_array(weights, init.name)
            new_initializers.append(new_initializer)
        else:
            new_initializers.append(init)

    # 创建新的图
    new_graph = onnx.helper.make_graph(
        nodes=model.graph.node,
        name=model.graph.name,
        inputs=model.graph.input,
        outputs=model.graph.output,
        initializer=new_initializers
    )

    # 创建新的模型
    new_model = onnx.helper.make_model(new_graph, producer_name='onnx-example')
    # 保存剪枝后的模型
    onnx.save(new_model, pruned_model)


def find_onnx_files(root_dir):
    onnx_files = []
    # 遍历根目录及其子目录
    for dirpath, _, filenames in os.walk(root_dir):
        # 查找所有以 .onnx 结尾的文件
        for filename in filenames:
            if filename.endswith('.onnx'):
                # 获取完整路径并添加到列表
                onnx_files.append(os.path.join(dirpath, filename))
    return onnx_files


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='模型文件剪枝工具')
    parser.add_argument('--target_dir', default=None, type=str, help='待剪枝的模型文件存放根目录，支持子文件夹递归处理')
    parser.add_argument('--pruned_saved_dir', default=None, type=str, help='剪枝模型文件保存目录，默认为None，表示与原始onnx权重文件放在同一目录下')
    parser.add_argument('--percent', default=0.05, type=float, help='规则剪枝百分比')
    args, _ = parser.parse_known_args()
    if args.target_dir is None:
        raise Exception("模型目录参数不可为空")

    onnx_files = find_onnx_files(args.target_dir)
    for onnx_file in onnx_files:
        if args.pruned_saved_dir:
            pruned_file = args.pruned_saved_dir + '/' + os.path.basename(onnx_file).replace('.onnx', '_pruned.onnx')
        else:
            pruned_file = onnx_file.replace('.onnx', '_pruned.onnx')
        prune_weights(model_path=onnx_file, pruned_model=pruned_file, pruning_percentage=args.percent)