pytorch多标签分类代码

转载

mob64ca141677f9 2024-09-13 20:09:15

文章标签 pytorch多标签分类代码 ML python P4 文章分类 PyTorch 人工智能

背景：之前的网络结构做多标签分类较为复杂，我们需要用resNet进行多标签分类简单的实现相应的分类问题，测试baseline。

一、网络结构及定义

二、optimizer

三、新定义网络

3.1 加载模型

3.2 模型定义

3.3 网络尺寸

3.4 fc层尺寸

四、训练内存占用

五、SENet更改

5.1 引入模型

5.2 SENet定义代码

一、网络结构及定义

在原来代码基础上更改，方便运行。

加入resnet101的结构

elif Config.MODEL == 'resnet101':
        model = models.resnet101(pretrained=False)
        print('load pretrained model...')
        model.load_state_dict(torch.load('./resnet101-5d3b4d8f.pth'))

# model params
    MODEL = 'resnet101'  # options: hgat_conv, hgat_fc, groupnet
    BACKBONE = 'resnet101'
    GROUPS = 12

二、optimizer

报错：

optimizer = torch.optim.SGD(model.parameters(), 'lr':parser.lr,
                                                        ^
SyntaxError: invalid syntax

之前的optimizer

optimizer = torch.optim.SGD(model.get_config_optim(args.lr, args.lrp),
                                    # lr=args.lr,
                                    momentum=args.momentum,
                                    weight_decay=args.weight_decay)


    def get_config_optim(self, lr, lrp):
        return [
            {'params': self.features.parameters(), 'lr': lrp},
            {'params': self.heads.parameters(), 'lr': lr},
        ]

加入resnet之后，相应的optimizer不能用之前的了。因为没有了相应的get_config_optim

if Config.MODEL == 'resnet101':
        optimizer = torch.optim.SGD(params=model.parameters(), lr=args.lr,
                                    # lr=args.lr,
                                    momentum=args.momentum,
                                    weight_decay=args.weight_decay)
    else:
        optimizer = torch.optim.SGD(model.get_config_optim(args.lr, args.lrp),
                                    # lr=args.lr,
                                    momentum=args.momentum,
                                    weight_decay=args.weight_decay)

绕进去之后，发现一层一层调试下去，最终的resnet输入与之前的不一样

File "/home/xingxiangrui/chun-ML_GCN/engine.py", line 256, in train
    self.on_forward(True, model, criterion, data_loader, optimizer)
  File "/home/xingxiangrui/chun-ML_GCN/engine.py", line 446, in on_forward
    self.state['output'] = model(feature_var, inp_var)
  File "/home/xingxiangrui/chun-ML_GCN/env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/xingxiangrui/chun-ML_GCN/env/lib/python3.6/site-packages/torch/nn/parallel/data_parallel.py", line 73, in forward
    outputs = self.parallel_apply(replicas, inputs, kwargs)
。。。
  File "/home/xingxiangrui/chun-ML_GCN/env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 357, in __call__
    result = self.forward(*input, **kwargs)
TypeError: forward() takes 2 positional arguments but 3 were given

所以，改为采用新网络，重新定义ml_resnet.py函数，然后import进来。

三、新定义网络

3.1 加载模型

elif Config.MODEL == 'resnet101':
        import mymodels.ml_resnet as ml_resnet
        model= ml_resnet.ML_RESNET(Config.BACKBONE, groups=Config.GROUPS, nclasses=Config.NCLASSES,
                            nclasses_per_group=Config.NCLASSES_PER_GROUP,
                            group_channels=Config.GROUP_CHANNELS, class_channels=Config.CLASS_CHANNELS)

3.2 模型定义

从文件中加载定义，创建新文件

class ML_RESNET(nn.Module):
    def __init__(self, backbone, groups, nclasses, nclasses_per_group, group_channels, class_channels):
        super(HGAT_FC, self).__init__()
        self.groups = groups
        self.nclasses = nclasses
        self.nclasses_per_group = nclasses_per_group
        self.group_channels = group_channels
        self.class_channels = class_channels
        if backbone == 'resnet101':
            model = models.resnet101(pretrained=False)
            print('load pretrained model...')
            model.load_state_dict(torch.load('./resnet101-5d3b4d8f.pth'))
        elif backbone == 'resnet50':
            model = models.resnet50(pretrained=False)
            print('load pretrained model...')
            model.load_state_dict(torch.load('./resnet50-5d3b4d8f.pth'))
        else:
            raise Exception()
        self.features = nn.Sequential(
            model.conv1,
            model.bn1,
            model.relu,
            model.maxpool,
            model.layer1,
            model.layer2,
            model.layer3,
            model.layer4, )
        self.gmp = nn.AdaptiveMaxPool2d(1)
        self.fc=nn.Sequential(utils.BasicLinear(in_channels=2048, out_channels=1024),
                                       utils.BasicLinear(in_channels=1024, out_channels=class_channels), )

        self.image_normalization_mean = [0.485, 0.456, 0.406]
        self.image_normalization_std = [0.229, 0.224, 0.225]

    def forward(self, x, inp):
        x = self.features(x)  # [B,2048,H,W]
        x=self.gmp(x).view(x.size(0),x.size(1))
        x=self.fc(x)
        return x

    def get_config_optim(self, lr, lrp):
        return [
            {'params': self.features.parameters(), 'lr': lrp},
            #{'params': self.heads.parameters(), 'lr': lr},
        ]

定义需要改成自己的定义，ML_RESNET

class ML_RESNET(nn.Module):
    def __init__(self, backbone, groups, nclasses, nclasses_per_group, group_channels, class_channels):
        super(ML_RESNET, self).__init__()

3.3 网络尺寸

之前的网络尺寸定义为：

raise Exception()
        self.features = nn.Sequential(
            model.conv1,
            model.bn1,
            model.relu,
            model.maxpool,
            model.layer1,
            model.layer2,
            model.layer3,
            model.layer4, )
        self.gmp = nn.AdaptiveMaxPool2d(1)
        self.fc=nn.Sequential(utils.BasicLinear(in_channels=2048, out_channels=1024),
                                       utils.BasicLinear(in_channels=1024, out_channels=class_channels), )

def forward(self, x, inp):
        x = self.features(x)  # [B,2048,H,W]
        x=self.gmp(x).view(x.size(0),x.size(1))
        x=self.fc(x)
        return x

结果产生报错

File "general_train.py", line 182, in <module>
    main_coco()
  File "general_train.py", line 178, in main_coco
    engine.learning(model, criterion, train_dataset, val_dataset, optimizer)
  File "/home/xingxiangrui/chun-ML_GCN/engine.py", line 214, in learning
    self.train(train_loader, model, criterion, optimizer, epoch)
  File "/home/xingxiangrui/chun-ML_GCN/engine.py", line 256, in train
    self.on_forward(True, model, criterion, data_loader, optimizer)
  File "/home/xingxiangrui/chun-ML_GCN/engine.py", line 451, in on_forward
    weight=torch.autograd.Variable(weights.cuda()))
  File "/home/xingxiangrui/chun-ML_GCN/env/lib/python3.6/site-packages/torch/nn/functional.py", line 1227, in binary_cross_entropy_with_logits
    raise ValueError("Target size ({}) must be the same as input size ({})".format(target.size(), input.size()))
ValueError: Target size (torch.Size([32, 80])) must be the same as input size (torch.Size([32, 256]))

出现这个错误的原因，是这样，，一个为80，一个为256

看报错，是交叉熵的过程中出现的错误

NCLASSES = 80
    NCLASSES_PER_GROUP = [1, 8, 5, 10, 5, 10, 7, 10, 6, 6, 5, 7]  # FIXME: to check
    GROUP_CHANNELS = 512
    CLASS_CHANNELS = 256

class channels与class numbers出现了错误，其中一个为256一个为80

定义时进行更改

self.features = nn.Sequential(
            model.conv1,
            model.bn1,
            model.relu,
            model.maxpool,
            model.layer1,
            model.layer2,
            model.layer3,
            model.layer4, )
        self.gmp = nn.AdaptiveMaxPool2d(1)
        self.fc=nn.Sequential(utils.BasicLinear(in_channels=2048, out_channels=1024),
                                       utils.BasicLinear(in_channels=1024, out_channels=nclasses), )

将最终输出结果改为nclasses

3.4 fc层尺寸

之前fc层选用错了，不该用basicLinear

self.fc=nn.Sequential(utils.BasicLinear(in_channels=2048, out_channels=1024),
                                       utils.BasicLinear(in_channels=1024, out_channels=nclasses), )

basiclinear是我们自己定义的。

class BasicLinear(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(BasicLinear, self).__init__()
        self.fc = nn.Linear(in_features=in_channels, out_features=out_channels, bias=False)
        self.bn = nn.BatchNorm1d(num_features=out_channels)
        self.relu = nn.ReLU()

加了ReLU可能导致无法收敛，我们应当直接加linear层来保证模型收敛。

更改self.fc尺寸

# self.fc=nn.Sequential(utils.BasicLinear(in_channels=2048, out_channels=1024),
        #                                utils.BasicLinear(in_channels=1024, out_channels=nclasses), )
        self.fc = nn.Linear(in_features=2048, out_features=nclasses, bias=True)

四、训练内存占用

训练过程依然占据四张显卡。且占用内存基本不变。

若用三张显卡训练，会报错，list out of memroy

[xingxiangrui@gzbh-mms-gpu55.gzbh.baidu.com chun-ML_GCN]$ nvidia-smi
Sun May  5 16:38:47 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 384.81                 Driver Version: 384.81                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla P4            On   | 00000000:02:00.0 Off |                    0 |
| N/A   61C    P0    64W /  75W |   5711MiB /  7606MiB |     47%      Default |
+-------------------------------+----------------------+----------------------+
|   1  Tesla P4            On   | 00000000:03:00.0 Off |                    0 |
| N/A   62C    P0    62W /  75W |   5349MiB /  7606MiB |     71%      Default |
+-------------------------------+----------------------+----------------------+
|   2  Tesla P4            On   | 00000000:82:00.0 Off |                    0 |
| N/A   58C    P0    65W /  75W |   5375MiB /  7606MiB |     79%      Default |
+-------------------------------+----------------------+----------------------+
|   3  Tesla P4            On   | 00000000:83:00.0 Off |                    0 |
| N/A   59C    P0    63W /  75W |   5351MiB /  7606MiB |     38%      Default |
+-------------------------------+----------------------+----------------------+

可以用watch -n 1 nvidia-smi指令，查看显卡动态占用。

五、SENet更改

背景：需要将模型之中加入SENet的内容，因此需要更改骨架。

SENet的PyTorch代码，此代码没有预训练：https://github.com/moskomule/senet.pytorch

加入新的有预训练的代码：

5.1 引入模型

从模型引入相应的class

import models.senet_origin as senet_origin

文件代码地址见：SENet的PyTorch代码：https://github.com/moskomule/senet.pytorch

带有预训练的SENet的PyTorch代码：https://github.com/Xingxiangrui/various_pyTorch_network_structure/blob/master/senet_and_pretrained.py

# fixme new SE-resnet backbone
        if backbone == 'resnet101':
            model = senet_origin.se_resnet101()
        elif backbone == 'resnet50':
            model = senet_origin.se_resnet50()
        elif backbone == 'resnet101_cbam':
            import mymodels.cbam as cbam
            model = cbam.resnet101_cbam()
        elif backbone=='resnet150':
            model=senet_origin.se_resnet152()
        else:
            raise Exception()
        # self.features = nn.Sequential(
        #     model.conv1,
        #     model.bn1,
        #     model.relu,
        #     model.maxpool,
        #     model.layer1,
        #     model.layer2,
        #     model.layer3,
        #     model.layer4, )
        self.features = nn.Sequential(
            # model.layer0,
            model.layer0,
            model.layer1,
            model.layer2,
            model.layer3,
            model.layer4 )