【深度学习】【pytorch】对卷积层置零卷积核进行真实剪枝-Toy模板网

这篇具有很好参考价值的文章主要介绍了【深度学习】【pytorch】对卷积层置零卷积核进行真实剪枝。希望对大家有所帮助。如果存在错误或未考虑完全的地方，请大家不吝赐教，您也可以点击"举报违法"按钮提交疑问。

最近需要对深度学习模型进行部署,因此需要对模型进行压缩,博主取舍了很多大佬的博文并亲测有效,分享笔记邀大家共同学习讨论

前言

深度学习剪枝(Pruning)是一种用于减少神经网络模型大小、减少计算量和提高推理效率的技术，通过去除神经网络中的冗余连接(权重)或节点(神经元)，从而实现模型的稀疏化。
深度学习剪枝(Pruning)具有以下几个好处：1. 模型压缩和存储节省；2. 计算资源节省；3. 加速推理速度；4. 防止过拟合。
“假剪枝”(Fake Pruning)是一种剪枝算法的称呼，它在剪枝过程中并不真正删除权重或节点，而是通过一些技巧将它们置零或禁用，以模拟剪枝的效果，不少优秀的论文就采用了"假剪枝"策略，尽管可以在一定程度上提高模型的推理速度，但假剪枝算法没有真正减少模型的大小，博主将通过讲解一个小案例，简洁易懂的说明一种对"假剪枝"卷积层进行真正的剪枝的的方法。

卷积层剪枝

可以先将最后的完整代码拷贝到自己的py文件中，然后按照博主的思路学习如何将置零卷积核进行真实剪枝：

初始化卷积层，并查看卷积层权重

# 示例使用一个具有3个输入通道和5个输出通道的卷积层
conv = nn.Conv2d(3, 5, 3)
print("原始卷积层权重：")
print(conv.weight.data)
print(conv.weight.size())
print("原始卷积层偏置：")
print(conv.bias.data)
print(conv.bias.size())

通过随机函数让部分卷积核权重置为0，模拟完成了假剪枝。

# remove_zero_kernels方法内的代码
weight = conv_layer.weight.data
# 卷积核个数
num_kernels = weight.size(0)
# 随机对部分卷积置0
pruned = torch.ones(num_kernels, 1, 1, 1)
# 选择随着置0的卷积序号
random_int = random.randint(1, num_kernels-1)
for i in range(random_int):
    pruned[i, 0, 0, 0] = 0
conv_layer.weight.data = weight * pruned
weight = conv_layer.weight.data
bias = conv_layer.bias.data

保存未被剪枝的卷积核的权重和偏置

# 计算每个卷积核的L2范数,目的是为了检查卷积核的所有位置是不是都置0了
norms = torch.norm(weight.view(num_kernels, -1), dim=1)
zero_kernel_indices = torch.nonzero(norms==0).squeeze()
print(zero_kernel_indices)
# 移除L2范数为零的卷积核
new_weight = torch.stack([weight[i, :, :, :] for i in range(num_kernels) if i not in zero_kernel_indices])
new_bias = torch.stack([bias[i] for i in range(num_kernels) if i not in zero_kernel_indices])

构建新的卷积层，用来替换此前的卷积层，完成置零卷积核的真实剪枝

# 构建新的卷积层
if zero_kernel_indices.numel() > 0:
    # 输入channel
    in_channels = weight.size(1)
    # 输出channel
    out_channels = new_weight.size(0)
    # 卷积核大小
    kernel_size = weight.size(2)
    # 步长
    stride = conv_layer.stride
    padding = conv_layer.padding
    dilation = conv_layer.dilation
    groups = conv_layer.groups
    new_conv_layer = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, groups)
    new_conv_layer.weight.data = new_weight
    new_conv_layer.bias.data = new_bias
else:
    new_conv_layer = conv_layer

完整代码

import torch
import torch.nn as nn
import random

def remove_zero_kernels(conv_layer):
    # 卷积核权重
    weight = conv_layer.weight.data
    # 卷积核个数
    num_kernels = weight.size(0)
    # 随机对部分卷积置0
    pruned = torch.ones(num_kernels, 1, 1, 1)
    # 选择随着置0的卷积序号
    random_int = random.randint(1, num_kernels-1)
    for i in range(random_int):
        pruned[i, 0, 0, 0] = 0
    conv_layer.weight.data = weight * pruned
    weight = conv_layer.weight.data
    bias = conv_layer.bias.data
    # 计算每个卷积核的L2范数，目的是为了检查卷积核的所有位置是不是都置0了
    norms = torch.norm(weight.view(num_kernels, -1), dim=1)
    zero_kernel_indices = torch.nonzero(norms==0).squeeze()
    print(zero_kernel_indices)
    # 移除L2范数为零的卷积核
    new_weight = torch.stack([weight[i, :, :, :] for i in range(num_kernels) if i not in zero_kernel_indices])
    new_bias = torch.stack([bias[i] for i in range(num_kernels) if i not in zero_kernel_indices])
    # 构建新的卷积层
    if zero_kernel_indices.numel() > 0:
        # 输入channel
        in_channels = weight.size(1)
        # 输出channel
        out_channels = new_weight.size(0)
        # 卷积核大小
        kernel_size = weight.size(2)
        # 步长
        stride = conv_layer.stride
        padding = conv_layer.padding
        dilation = conv_layer.dilation
        groups = conv_layer.groups
        new_conv_layer = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, groups)
        new_conv_layer.weight.data = new_weight
        new_conv_layer.bias.data = new_bias
    else:
        new_conv_layer = conv_layer

    return new_conv_layer

# 示例使用一个具有3个输入通道和5个输出通道的卷积层
conv = nn.Conv2d(3, 5, 3)
# print("原始卷积层权重：")
# print(conv.weight.data)
# print(conv.weight.size())
# print("原始卷积层偏置：")
# print(conv.bias.data)
# print(conv.bias.size())

# 将置零的卷积核移除
new_conv = remove_zero_kernels(conv)
# print("原始卷积层权重：")
# print(new_conv.weight.data)
# print(new_conv.weight.size())
# print("原始卷积层偏置：")
# print(new_conv.bias.data)
# print(new_conv.bias.size())