模型剪枝大瘦身代码实战案例

模型剪枝大瘦身代码实战案例

注：大家觉得博客好的话，别忘了点赞收藏呀，本人每周都会更新关于人工智能和大数据相关的内容，内容多为原创，Python Java Scala SQL 代码，CV NLP 推荐系统等，Spark Flink Kafka Hbase Hive Flume等等~写的都是纯干货，各种顶会的论文解读，一起进步。
今天和大家分享一下如何通过BN加上L1正则化的方法对模型进行剪枝大瘦身
Learning Efficient Convolutional Networks through Network Slimming
https://github.com/foolwood/pytorch-slimming
#博学谷IT学习技术支持#

---

---

前言

今天和大家分享一篇ICCV的经典论文，通俗易懂，非常简单，主要是通过对BN中gamma加上L1正则化的方法对模型进行剪枝大瘦身。

---

一、损失函数

现有的一些稀疏化方法，包括权重稀疏化、核稀疏化、通道稀疏化和层稀疏化等。权重稀疏化灵活性最高，也获得最高的压缩率，但是通常需要特定硬件的支持才能达到加速的效果。层稀疏化最粗糙，灵活性最低，但是在一般的硬件上都能获得加速效果。通道裁剪有一个比较好的均衡，而且也可以在包括 CNN 或者全连层网络上都可以进行应用。

为了达到通道稀疏性的效果，和这个通道相关的输入和输出连接都需要被裁剪。

论文提出的方法，是对每个通道训练一个尺度因子，将这个尺度因子和通道输出进行相乘。在网络训练时对网络和这些尺度因子进行联合优化，然后裁剪掉尺度因子小的通道，并对裁剪后的网络进行微调。因此网络的损失函数就是：

这里和一般的损失函数相比唯一的区别就是对BN中的参数gamma加了L1的正则项。这样通过L1的正则项，可以使得那些原本权重偏小的gamma更趋向于0。

二、整体框架

通过BN中正则化以后gamma值的差异，可以看出CNN中哪写特征图是对模型更有用的。留下这些特征图即可。

然后在对模型进行微调重新训练即可。模型效果不会差太多。

三、代码实战

1.训练模型

这里和一般训练模型的代码没什么区别，唯一不一样的是
定义updateBN函数，因为要对gamma加上L1正则项。
def updateBN():
for m in model.modules():
if isinstance(m, nn.BatchNorm2d):
m.weight.grad.data.add_(args.s*torch.sign(m.weight.data)) # L1 大于0为1 小于0

代码如下（示例）：

from __future__ import print_function
import os
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
from torch.autograd import Variable

from vgg import vgg
import shutil

# 1：训练，并且加入l1正则化 -sr --s 0.0001
# 2：执行剪枝操作 --model model_best.pth.tar --save pruned.pth.tar --percent 0.7
# 3：再次进行微调操作 --refine pruned.pth.tar --epochs 40

# Training settings
parser = argparse.ArgumentParser(description='PyTorch Slimming CIFAR training')
parser.add_argument('--dataset', type=str, default='cifar10',
                    help='training dataset (default: cifar10)')
parser.add_argument('--sparsity-regularization', '-sr', dest='sr', action='store_true',
                    help='train with channel sparsity regularization')
parser.add_argument('--s', type=float, default=0.0001,
                    help='scale sparse rate (default: 0.0001)')
parser.add_argument('--refine', default='', type=str, metavar='PATH',
                    help='refine from prune model')
parser.add_argument('--batch-size', type=int, default=100, metavar='N',
                    help='input batch size for training (default: 100)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                    help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=5, metavar='N',
                    help='number of epochs to train (default: 160)')
parser.add_argument('--start-epoch', default=0, type=int, metavar='N',
                    help='manual epoch number (useful on restarts)')
parser.add_argument('--lr', type=float, default=0.1, metavar='LR',
                    help='learning rate (default: 0.1)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                    help='SGD momentum (default: 0.9)')
parser.add_argument('--weight-decay', '--wd', default=1e-4, type=float,
                    metavar='W', help='weight decay (default: 1e-4)')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
                    help='path to latest checkpoint (default: none)')
parser.add_argument('--no-cuda', action='store_true', default=False,
                    help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
                    help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=100, metavar='N',
                    help='how many batches to wait before logging training status')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()

torch.manual_seed(args.seed)
if args.cuda:
    torch.cuda.manual_seed(args.seed)


kwargs = {'num_workers': 0, 'pin_memory': True} if args.cuda else {}
train_loader = torch.utils.data.DataLoader(
    datasets.CIFAR10('./data', train=True, download=True,
                   transform=transforms.Compose([
                       transforms.Pad(4),
                       transforms.RandomCrop(32),
                       transforms.RandomHorizontalFlip(),
                       transforms.ToTensor(),
                       transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
                   ])),
    batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
    datasets.CIFAR10('./data', train=False, transform=transforms.Compose([
                       transforms.ToTensor(),
                       transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
                   ])),
    batch_size=args.test_batch_size, shuffle=True, **kwargs)

if args.refine:
    checkpoint = torch.load(args.refine)
    model = vgg(cfg=checkpoint['cfg'])
    model.cuda()
    model.load_state_dict(checkpoint['state_dict'])
else:
    model = vgg()
if args.cuda:
    model.cuda()

optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)

if args.resume:
    if os.path.isfile(args.resume):
        print("=> loading checkpoint '{}'".format(args.resume))
        checkpoint = torch.load(args.resume)
        args.start_epoch = checkpoint['epoch']
        best_prec1 = checkpoint['best_prec1']
        model.load_state_dict(checkpoint['state_dict'])
        optimizer.load_state_dict(checkpoint['optimizer'])
        print("=> loaded checkpoint '{}' (epoch {}) Prec1: {:f}"
              .format(args.resume, checkpoint['epoch'], best_prec1))
    else:
        print("=> no checkpoint found at '{}'".format(args.resume))

# additional subgradient descent on the sparsity-induced penalty term
def updateBN():
    for m in model.modules():
        if isinstance(m, nn.BatchNorm2d):
            m.weight.grad.data.add_(args.s*torch.sign(m.weight.data))  # L1 大于0为1 小于0为-1 0还是0


def train(epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        if args.cuda:
            data, target = data.cuda(), target.cuda()
        data, target = Variable(data), Variable(target)
        optimizer.zero_grad()
        output = model(data)
        loss = F.cross_entropy(output, target)
        loss.backward()
        if args.sr:
            updateBN()
        optimizer.step()
        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.1f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))

def test():
    model.eval()
    test_loss = 0
    correct = 0
    for data, target in test_loader:
        if args.cuda:
            data, target = data.cuda(), target.cuda()
        data, target = Variable(data, volatile=True), Variable(target)
        output = model(data)
        test_loss += F.cross_entropy(output, target, size_average=False).item() # sum up batch loss
        pred = output.data.max(1, keepdim=True)[1] # get the index of the max log-probability
        correct += pred.eq(target.data.view_as(pred)).cpu().sum()

    test_loss /= len(test_loader.dataset)
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.1f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        100. * correct / len(test_loader.dataset)))
    return correct / float(len(test_loader.dataset))


def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'):
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, 'model_best.pth.tar')

best_prec1 = 0.
for epoch in range(args.start_epoch, args.epochs):
    if epoch in [args.epochs*0.5, args.epochs*0.75]:
        for param_group in optimizer.param_groups:
            param_group['lr'] *= 0.1
    train(epoch)
    prec1 = test()
    is_best = prec1 > best_prec1
    best_prec1 = max(prec1, best_prec1)
    save_checkpoint({
        'epoch': epoch + 1,
        'state_dict': model.state_dict(),
        'best_prec1': best_prec1,
        'optimizer': optimizer.state_dict(),
    }, is_best)

2.模型剪纸瘦身

这里通过gamma阈值剪去无用的特征图后，对模型进行再训练微调，使模型整体效果得到提升。
代码如下（示例）：

import os
import argparse
import torch
import torch.nn as nn
from torch.autograd import Variable
from torchvision import datasets, transforms

from vgg import vgg
import numpy as np

# Prune settings
parser = argparse.ArgumentParser(description='PyTorch Slimming CIFAR prune')
parser.add_argument('--dataset', type=str, default='cifar10',
                    help='training dataset (default: cifar10)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                    help='input batch size for testing (default: 1000)')
parser.add_argument('--no-cuda', action='store_true', default=False,
                    help='disables CUDA training')
parser.add_argument('--percent', type=float, default=0.5,
                    help='scale sparse rate (default: 0.5)')
parser.add_argument('--model', default='', type=str, metavar='PATH',
                    help='path to raw trained model (default: none)')
parser.add_argument('--save', default='', type=str, metavar='PATH',
                    help='path to save prune model (default: none)')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()

model = vgg()
if args.cuda:
    model.cuda()
if args.model:
    if os.path.isfile(args.model):
        print("=> loading checkpoint '{}'".format(args.model))
        checkpoint = torch.load(args.model)
        args.start_epoch = checkpoint['epoch']
        best_prec1 = checkpoint['best_prec1']
        model.load_state_dict(checkpoint['state_dict'])
        print("=> loaded checkpoint '{}' (epoch {}) Prec1: {:f}"
              .format(args.model, checkpoint['epoch'], best_prec1))
    else:
        print("=> no checkpoint found at '{}'".format(args.resume))

print(model)
total = 0 # 每层特征图个数 总和
for m in model.modules():
    if isinstance(m, nn.BatchNorm2d):
        total += m.weight.data.shape[0]

bn = torch.zeros(total) # 拿到每一个gamma值 每个特征图都会对应一个γ、β
index = 0
for m in model.modules():
    if isinstance(m, nn.BatchNorm2d):
        size = m.weight.data.shape[0]
        bn[index:(index+size)] = m.weight.data.abs().clone()
        index += size

y, i = torch.sort(bn)
thre_index = int(total * args.percent)
thre = y[thre_index]

pruned = 0
cfg = []
cfg_mask = []
for k, m in enumerate(model.modules()):
    if isinstance(m, nn.BatchNorm2d):
        weight_copy = m.weight.data.clone()
        mask = weight_copy.abs().gt(thre).float().cuda() #.gt 比较前者是否大于后者
        pruned = pruned + mask.shape[0] - torch.sum(mask)
        m.weight.data.mul_(mask) # BN层gamma置0
        m.bias.data.mul_(mask) #
        cfg.append(int(torch.sum(mask)))
        cfg_mask.append(mask.clone())
        print('layer index: {:d} \t total channel: {:d} \t remaining channel: {:d}'.
            format(k, mask.shape[0], int(torch.sum(mask))))
    elif isinstance(m, nn.MaxPool2d):
        cfg.append('M')

pruned_ratio = pruned/total

print('Pre-processing Successful!')


# 置0后先测试下效果
def test():
    kwargs = {'num_workers': 0, 'pin_memory': True} if args.cuda else {}
    test_loader = torch.utils.data.DataLoader(
        datasets.CIFAR10('./data', train=False, transform=transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])),
        batch_size=args.test_batch_size, shuffle=True, **kwargs)
    model.eval()
    correct = 0
    for data, target in test_loader:
        if args.cuda:
            data, target = data.cuda(), target.cuda()
        data, target = Variable(data, volatile=True), Variable(target)
        output = model(data)
        pred = output.data.max(1, keepdim=True)[1] # get the index of the max log-probability
        correct += pred.eq(target.data.view_as(pred)).cpu().sum()

    print('\nTest set: Accuracy: {}/{} ({:.1f}%)\n'.format(
        correct, len(test_loader.dataset), 100. * correct / len(test_loader.dataset)))
    return correct / float(len(test_loader.dataset))

# test()


# 执行剪枝
print(cfg)
newmodel = vgg(cfg=cfg) # 剪枝后的模型
newmodel.cuda()
# 为剪枝后的模型赋值权重
layer_id_in_cfg = 0
start_mask = torch.ones(3) #输入
end_mask = cfg_mask[layer_id_in_cfg] #输出
for [m0, m1] in zip(model.modules(), newmodel.modules()):
    if isinstance(m0, nn.BatchNorm2d): 
        idx1 = np.squeeze(np.argwhere(np.asarray(end_mask.cpu().numpy()))) # 赋值
        m1.weight.data = m0.weight.data[idx1].clone()
        m1.bias.data = m0.bias.data[idx1].clone()
        m1.running_mean = m0.running_mean[idx1].clone()
        m1.running_var = m0.running_var[idx1].clone()
        layer_id_in_cfg += 1
        start_mask = end_mask.clone() #下一层的
        if layer_id_in_cfg < len(cfg_mask):  # do not change in Final FC
            end_mask = cfg_mask[layer_id_in_cfg] #输出
    elif isinstance(m0, nn.Conv2d):
        idx0 = np.squeeze(np.argwhere(np.asarray(start_mask.cpu().numpy())))
        idx1 = np.squeeze(np.argwhere(np.asarray(end_mask.cpu().numpy())))
        print('In shape: {:d} Out shape:{:d}'.format(idx0.shape[0], idx1.shape[0]))
        w = m0.weight.data[:, idx0, :, :].clone() #拿到原始训练好权重
        w = w[idx1, :, :, :].clone()
        m1.weight.data = w.clone() # 将所需权重赋值到剪枝后的模型
        # m1.bias.data = m0.bias.data[idx1].clone()
    elif isinstance(m0, nn.Linear):
        idx0 = np.squeeze(np.argwhere(np.asarray(start_mask.cpu().numpy())))
        m1.weight.data = m0.weight.data[:, idx0].clone()


torch.save({'cfg': cfg, 'state_dict': newmodel.state_dict()}, args.save)

print(newmodel)
model = newmodel
test()

3.VGG模型本身

不多说了，一个比较老比较经典的模型举个例子，大家可以根据自己的下游任务替换即可。
代码如下（示例）：

import torch
import torch.nn as nn
from torch.autograd import Variable
import math  # init


class vgg(nn.Module):

    def __init__(self, dataset='cifar10', init_weights=True, cfg=None):
        super(vgg, self).__init__()
        if cfg is None:
            cfg = [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512]
        self.feature = self.make_layers(cfg, True)

        if dataset == 'cifar100':
            num_classes = 100
        elif dataset == 'cifar10':
            num_classes = 10
        self.classifier = nn.Linear(cfg[-1], num_classes)
        if init_weights:
            self._initialize_weights()

    def make_layers(self, cfg, batch_norm=False):
        layers = []
        in_channels = 3
        for v in cfg:
            if v == 'M':
                layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
            else:
                conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1, bias=False)
                #print (in_channels,'   ',v)
                if batch_norm:
                    layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)]
                else:
                    layers += [conv2d, nn.ReLU(inplace=True)]
                in_channels = v
        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.feature(x)
        x = nn.AvgPool2d(2)(x)
        x = x.view(x.size(0), -1)
        y = self.classifier(x)
        return y

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(0.5)
                m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                m.weight.data.normal_(0, 0.01)
                m.bias.data.zero_()


if __name__ == '__main__':
    net = vgg()
    x = Variable(torch.FloatTensor(16, 3, 40, 40))
    y = net(x)
    print(y.data.shape)

---

总结

非常直观的模型剪枝大瘦身方法非常好用，文章的核心在于，将BN层中的可学习参数gamma，作为稀疏化参数，对CNN卷积之后提取保留更有用的特征图的方法进行模型剪枝大瘦身。