PyTorch–模型剪枝案例
一、基础知识：
1.模型剪枝：
通俗理解就是将神经网络某些冗余连接层的权重置为0，使得模型更加具有稀疏化，从而提升模型性能
下图通过掩码图，根据掩码图对应权重矩阵将对应位置上的值替换为0。
2.模型剪枝的方式：
剪枝可以分为非结构剪枝和结构剪枝：
非结构剪枝：
如下图（左）：神经元的个数不变，将某些神经元的部分连接权重置为零（虚线部分）
结构剪枝：
如下图（右）：神经元的个数不变，将某些神经元的全部连接权重置为零（虚线部分）
二、pytorch中的官方模型剪枝：
1.局部剪枝
1.首先构建一个网络模型，根据torch官方采用LeNet：
'''
https://pytorch.org/tutorials/intermediate/pruning_tutorial.html
'''
import torch
from torch import nn
import torch.nn.utils.prune as prune
import torch.nn.functional as F
# 选定设备
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# 构建LeNet网络模型
class LeNet(nn.Module):
def __init__(self):
super(LeNet, self).__init__()
self.conv1 = nn.Conv2d(in_channels=1, out_channels=6, kernel_size=3)
self.conv2 = nn.Conv2d(6, 16, 3)
self.fc1 = nn.Linear(16 * 5 * 5, 120) # 5x5 image dimension
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10)
def forward(self, x):
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
x = x.view(-1, int(x.nelement() / x.shape[0]))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
model = LeNet().to(device=device) # 模型传入cpu/gpu
module = model.conv1 # 选定为第一个卷积层为操作的模块
查看第一层卷积的参数: 包含weight和bias
named_parameters()内存储的对象除非手动删除，否则在剪枝过程中对其无影响
print(list(module.named_parameters()))
[('weight', Parameter containing:
tensor([[[[ 1.5996e-01, 1.0775e-01, 2.6948e-01],
[-6.4919e-02, -1.8932e-01, 2.2985e-02],
[ 8.5653e-02, 2.9077e-01, 2.0981e-01]]],
[[[-2.0450e-01, 5.4691e-02, 2.4475e-01],
[ 3.1142e-01, 1.9070e-01, -1.7687e-01],
[-3.9149e-02, 3.1566e-01, 1.8814e-02]]],
[[[ 3.1525e-01, 6.9234e-03, 2.1742e-01],
[ 1.2825e-02, -1.4790e-01, 1.2469e-01],
[ 2.0399e-01, -1.1661e-01, 2.5957e-01]]],
[[[-8.3796e-02, -2.3485e-01, -2.3683e-01],
[-3.1888e-01, -2.5757e-01, 9.3788e-02],
[-2.6058e-01, -1.6827e-01, 3.0230e-02]]],
[[[-1.4831e-02, -2.8372e-04, -2.2037e-01],
[-2.6850e-01, -1.9531e-01, -3.0480e-01],
[ 2.1453e-02, 1.0219e-02, 1.0799e-01]]],
[[[ 6.6237e-03, -1.1856e-02, 2.1806e-01],
[-1.2661e-01, -1.2371e-01, 2.9971e-01],
[ 4.9645e-03, 7.6150e-02, -1.2875e-01]]]], device='cuda:0',
requires_grad=True)), ('bias', Parameter containing:
tensor([ 0.1880, -0.3062, -0.3171, -0.1479, 0.2673, -0.1545], device='cuda:0',
requires_grad=True))]
2.对weight和bias的剪枝
（1）weight剪枝：
# 选择需要剪枝的参数，随机非结构化剪掉30%的参数
prune.random_unstructured(module, name="weight", amount=0.3)
剪枝后查看named_parameters（不变）：
print(list(module.named_parameters())) # 查看第一层卷积层的参数，包括weight和bias
[('bias', Parameter containing:
tensor([ 0.1880, -0.3062, -0.3171, -0.1479, 0.2673, -0.1545], device='cuda:0',
requires_grad=True)), ('weight_orig', Parameter containing:
tensor([[[[ 1.5996e-01, 1.0775e-01, 2.6948e-01],
[-6.4919e-02, -1.8932e-01, 2.2985e-02],
[ 8.5653e-02, 2.9077e-01, 2.0981e-01]]],
[[[-2.0450e-01, 5.4691e-02, 2.4475e-01],
[ 3.1142e-01, 1.9070e-01, -1.7687e-01],
[-3.9149e-02, 3.1566e-01, 1.8814e-02]]],
[[[ 3.1525e-01, 6.9234e-03, 2.1742e-01],
[ 1.2825e-02, -1.4790e-01, 1.2469e-01],
[ 2.0399e-01, -1.1661e-01, 2.5957e-01]]],
[[[-8.3796e-02, -2.3485e-01, -2.3683e-01],
[-3.1888e-01, -2.5757e-01, 9.3788e-02],
[-2.6058e-01, -1.6827e-01, 3.0230e-02]]],
[[[-1.4831e-02, -2.8372e-04, -2.2037e-01],
[-2.6850e-01, -1.9531e-01, -3.0480e-01],
[ 2.1453e-02, 1.0219e-02, 1.0799e-01]]],
[[[ 6.6237e-03, -1.1856e-02, 2.1806e-01],
[-1.2661e-01, -1.2371e-01, 2.9971e-01],
[ 4.9645e-03, 7.6150e-02, -1.2875e-01]]]], device='cuda:0',
requires_grad=True))]
查看本次剪枝的掩码，所有剪枝操作的掩码都存储与named_buffers中（其中0和1分别代表丢弃和保留该wei位置的参数）：
print(list(module.named_buffers()))
[('weight_mask', tensor([[[[1., 1., 0.],
[1., 0., 0.],
[1., 0., 1.]]],
[[[1., 1., 1.],
[0., 0., 1.],
[0., 1., 1.]]],
[[[1., 0., 0.],
[1., 0., 1.],
[1., 1., 1.]]],
[[[1., 1., 1.],
[1., 1., 0.],
[1., 1., 1.]]],
[[[1., 0., 1.],
[1., 0., 0.],
[1., 1., 0.]]],
[[[1., 1., 1.],
[0., 1., 1.],
[1., 1., 1.]]]], device='cuda:0'))]
查看剪枝后权重参数（对应掩码位置为0的位置上的参数已经被替换为0）：
print(module.weight)
tensor([[[[ 0.1600, 0.1078, 0.0000],
[-0.0649, -0.0000, 0.0000],
[ 0.0857, 0.0000, 0.2098]]],
[[[-0.2045, 0.0547, 0.2447],
[ 0.0000, 0.0000, -0.1769],
[-0.0000, 0.3157, 0.0188]]],
[[[ 0.3153, 0.0000, 0.0000],
[ 0.0128, -0.0000, 0.1247],
[ 0.2040, -0.1166, 0.2596]]],
[[[-0.0838, -0.2349, -0.2368],
[-0.3189, -0.2576, 0.0000],
[-0.2606, -0.1683, 0.0302]]],
[[[-0.0148, -0.0000, -0.2204],
[-0.2685, -0.0000, -0.0000],
[ 0.0215, 0.0102, 0.0000]]],
[[[ 0.0066, -0.0119, 0.2181],
[-0.0000, -0.1237, 0.2997],
[ 0.0050, 0.0762, -0.1288]]]], device='cuda:0',
grad_fn=<MulBackward0>)
（2）对bias进行剪枝：
# 使用l1正则非结构剪枝对bias进行剪枝--->将绝对值最小的3个bias的值改为0
prune.l1_unstructured(module, name="bias", amount=3)
查看剪枝前后的bias
prune.l1_unstructured(module, name="bias", amount=3)
print("bias剪枝前:",list(module.named_parameters()))
print("bias剪枝后:",module.bias)
bias剪枝前: [('weight_orig', Parameter containing:
tensor([[[[ 1.5996e-01, 1.0775e-01, 2.6948e-01],
[-6.4919e-02, -1.8932e-01, 2.2985e-02],
[ 8.5653e-02, 2.9077e-01, 2.0981e-01]]],
[[[-2.0450e-01, 5.4691e-02, 2.4475e-01],
[ 3.1142e-01, 1.9070e-01, -1.7687e-01],
[-3.9149e-02, 3.1566e-01, 1.8814e-02]]],
[[[ 3.1525e-01, 6.9234e-03, 2.1742e-01],
[ 1.2825e-02, -1.4790e-01, 1.2469e-01],
[ 2.0399e-01, -1.1661e-01, 2.5957e-01]]],
[[[-8.3796e-02, -2.3485e-01, -2.3683e-01],
[-3.1888e-01, -2.5757e-01, 9.3788e-02],
[-2.6058e-01, -1.6827e-01, 3.0230e-02]]],
[[[-1.4831e-02, -2.8372e-04, -2.2037e-01],
[-2.6850e-01, -1.9531e-01, -3.0480e-01],
[ 2.1453e-02, 1.0219e-02, 1.0799e-01]]],
[[[ 6.6237e-03, -1.1856e-02, 2.1806e-01],
[-1.2661e-01, -1.2371e-01, 2.9971e-01],
[ 4.9645e-03, 7.6150e-02, -1.2875e-01]]]], device='cuda:0',
requires_grad=True)), ('bias_orig', Parameter containing:
tensor([ 0.1880, -0.3062, -0.3171, -0.1479, 0.2673, -0.1545], device='cuda:0',
requires_grad=True))]
bias剪枝后: tensor([ 0.0000, -0.3062, -0.3171, -0.0000, 0.2673, -0.0000], device='cuda:0',
grad_fn=<MulBackward0>)
（3）迭代剪枝：
# 此操作将剪掉2个node的weight,总的剪除数量为50%
prune.ln_structured(module, name="weight", amount=0.5, n=2, dim=0)
查看剪枝后的权重
print(module.weight)
tensor([[[[ 0.0000, 0.0000, 0.0000],
[-0.0000, -0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]]],
[[[-0.2045, 0.0547, 0.2447],
[ 0.0000, 0.0000, -0.1769],
[-0.0000, 0.3157, 0.0188]]],
[[[ 0.3153, 0.0000, 0.0000],
[ 0.0128, -0.0000, 0.1247],
[ 0.2040, -0.1166, 0.2596]]],
[[[-0.0838, -0.2349, -0.2368],
[-0.3189, -0.2576, 0.0000],
[-0.2606, -0.1683, 0.0302]]],
[[[-0.0000, -0.0000, -0.0000],
[-0.0000, -0.0000, -0.0000],
[ 0.0000, 0.0000, 0.0000]]],
[[[ 0.0000, -0.0000, 0.0000],
[-0.0000, -0.0000, 0.0000],
[ 0.0000, 0.0000, -0.0000]]]], device='cuda:0',
grad_fn=<MulBackward0>)
（4）对整个模型进行剪枝：
new_model = LeNet()
for name, module in new_model.named_modules():
# 在所有的2d卷积层中使用l1正则非结构剪枝20%的权重
if isinstance(module, torch.nn.Conv2d):
prune.l1_unstructured(module, name='weight', amount=0.2)
# 在所有的线性层使用l1正则非结构剪枝40%的权重
elif isinstance(module, torch.nn.Linear):
prune.l1_unstructured(module, name='weight', amount=0.4)
print(dict(new_model.named_buffers()).keys())
757
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
