J5、DenseNet+SE-Net实战
- 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
- 🍖 原作者：K同学啊|接辅导、项目定制
学习要求¶
- 在DenseNet系列算法中插入SE-Net通道注意力机制，并完成猴痘病识别
这里由于我换电脑了懒得下数据，使用了咖啡豆的图像
- 改进思路是否可以迁移到其他地方呢
- 测试集accuracy到达89%（拔高，可选）
一、背景介绍¶
SE-Net 是 ImageNet 2017（ImageNet 收官赛）的冠军模型，是由WMW团队发布。具有复杂度低，参数少和计算量小的优点。且SENet 思路很简单，很容易扩展到已有网络结构如 Inception 和 ResNet 中。 已经有很多工作在空间维度上来提升网络的性能，如 Inception 等，而 SENet 将关注点放在了特征通道之间的关系上。其具体策略为：通过学习的方式来自动获取到每个特征通道的重要程度，然后依照这个重要程度去提升有用的特征并抑制对当前任务用处不大的特征，这又叫做“特征重标定”策略。具体的 SE 模块如下图所示： 
1、首先是Squeeze操作，我们顺着空间维度来进行特征压缩，将一个通道数和输入的特征通道数相等，例如将形状为(1, 32, 32, 10)的feature map压缩成(1, 1, 1, 10)。此操作通常采用global average pooling来实现。 2、得到了全局描述特征后，进行Excitation操作来抓取特征通道之间的关系，它是一个类似于循环神经网络中门的机制。SE采用包含两个全连接层的bottleneck结构，即中间小两头大的结构：其中第一个全链接层起到即降维的作用，并通过ReLU激活，第二个全链接层用来将其恢复至原始的维度。进行Excitation操作的最终目的是为每个特征通道生成权重，即学习到的各个通道的激活值（sigmoid激活，值在0~1之间）。 3、最后一个是Scale的操作，我们将Excitation的输出的权重看作是经过特征选择后的每个特征通道的重要性，然后通过乘法逐通道加权到先前的特征上，完成在通道维度上的对原始特征的重标定，从而使得模型对各个通道的特征更有辨别能力，这类似于attention机制。
二、网络对比¶
SE模块的灵活性在于它可以直接应用在现有的网络结构中。以Inception和ResNet为例，我们只需要在Inception模块或Residual模块后添加一个SE模块即可。具体如下图所示： 
上图分别是将SE模块嵌入到Inception结构与ResNet中的示例，方框旁边的维度信息代表该层的输出， r rr 表示Excitation操作中的降维系数。
三、SE模块效果对比¶
SE模块很容易嵌入到其他网络中，为了验证SE模块的作用，在其它流行网络如ResNet和Inception中引入SE模块，测试其在ImageNet上的效果，如下表所示： 
首先看一下网络的深度对SE的影响。上表分别展示了ResNet-50、ResNet-101、ResNet-152、ResNeXt-50、ResNeXt-101和VGG-16、BN-Inception、Inception-ResNet-v2嵌入SE模型的结果。第一栏Original是原作者实现的结果，为了进行公平的比较，重新进行了实现实验得到re-implementation的结果。最后一栏SE-module是指嵌入了SE模块的结果，它的训练参数和第二栏re-implementation一致。括号中的红色数值是指相对于re-implementation的精度提升的幅值。
从上表可以看出，SE-ResNets在各种深度上都远远超过了其对应的没有SE的结构版本的精度，这说明无论网络的深度如何，SE模块都能够给网络带来性能上的增益。值得一提的是，SE-ResNet-50可以达到和ResNet-101一样的精度；更甚，SE-ResNet-101远远地超过了更深的ResNet-152。
四、SE模块代码实现¶
In [9]:
import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torchvision
import torch.nn.functional as F
from torchvision import transforms, datasets
from sklearn.model_selection import KFold
from torch.optim.lr_scheduler import StepLR, MultiStepLR, LambdaLR, ExponentialLR, CosineAnnealingLR, ReduceLROnPlateau
import os,PIL,pathlib,random
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
deviceOut[9]:
device(type='cuda')In [2]:
data_dir = './data/7-data'
# 通过Path类创建路径对象
data_dir = pathlib.Path(data_dir)
# 获取路径下所有文件路径
paths= list(data_dir.glob('*'))
# 获取所有文件夹的名字，也就是图片类别
classNames = [str(path).split("\\")[2] for path in paths] # K哥classNames中间会多一个e
classNamesOut[2]:
['Dark', 'Green', 'Light', 'Medium']In [3]:
train_transforms = transforms.Compose([
transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸
# transforms.RandomHorizontalFlip(), # 随机水平翻转
transforms.ToTensor(), # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
transforms.Normalize( # 标准化处理-->转换为标准正态分布（高斯分布），使模型更容易收敛
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])
test_transform = transforms.Compose([
transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸
transforms.ToTensor(), # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
transforms.Normalize( # 标准化处理-->转换为标准正态分布（高斯分布），使模型更容易收敛
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])
total_data = datasets.ImageFolder("./data/7-data/",transform=train_transforms)
total_dataOut[3]:
Dataset ImageFolder
Number of datapoints: 1200
Root location: ./data/7-data/
StandardTransform
Transform: Compose(
Resize(size=[224, 224], interpolation=bilinear, max_size=None, antialias=None)
ToTensor()
Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
)In [4]:
total_data.class_to_idxOut[4]:
{'Dark': 0, 'Green': 1, 'Light': 2, 'Medium': 3}In [5]:
train_size = int(0.8 * len(total_data))
test_size = len(total_data) - train_size
train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])
train_dataset, test_datasetOut[5]:
(<torch.utils.data.dataset.Subset at 0x20a3870fb80>,
<torch.utils.data.dataset.Subset at 0x20a3870f970>)In [6]:
batch_size = 8
train_dl = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=3)
test_dl = torch.utils.data.DataLoader(test_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=3)In [7]:
from collections import OrderedDict
import torch.utils.checkpoint as cp
def _bn_function_factory(norm, relu, conv):
def bn_function(*inputs):
concated_features = torch.cat(inputs, 1)
bottleneck_output = conv(relu(norm(concated_features)))
return bottleneck_output
return bn_function
class _DenseLayer(nn.Module):
def __init__(self, num_input_features, growth_rate, bn_size, drop_rate, efficient=False):
super(_DenseLayer, self).__init__()
self.add_module('norm1', nn.BatchNorm2d(num_input_features)),
self.add_module('relu1', nn.ReLU(inplace=True)),
self.add_module('conv1', nn.Conv2d(num_input_features, bn_size * growth_rate,
kernel_size=1, stride=1, bias=False)),
self.add_module('norm2', nn.BatchNorm2d(bn_size * growth_rate)),
self.add_module('relu2', nn.ReLU(inplace=True)),
self.add_module('conv2', nn.Conv2d(bn_size * growth_rate, growth_rate,
kernel_size=3, stride=1, padding=1, bias=False)),
self.add_module('SE_Block',SE_Block(growth_rate, reduction=16))
self.drop_rate = drop_rate
self.efficient = efficient
def forward(self, *prev_features):
bn_function = _bn_function_factory(self.norm1, self.relu1, self.conv1)
if self.efficient and any(prev_feature.requires_grad for prev_feature in prev_features):
bottleneck_output = cp.checkpoint(bn_function, *prev_features)
else:
bottleneck_output = bn_function(*prev_features)
new_features = self.SE_Block(self.conv2(self.relu2(self.norm2(bottleneck_output))))
if self.drop_rate > 0:
new_features = F.dropout(new_features, p=self.drop_rate, training=self.training)
return new_features
class _Transition(nn.Sequential):
def __init__(self, num_input_features, num_output_features):
super(_Transition, self).__init__()
self.add_module('norm', nn.BatchNorm2d(num_input_features))
self.add_module('relu', nn.ReLU(inplace=True))
self.add_module('conv', nn.Conv2d(num_input_features, num_output_features,
kernel_size=1, stride=1, bias=False))
self.add_module('pool', nn.AvgPool2d(kernel_size=2, stride=2))
class _DenseBlock(nn.Module):
def __init__(self, num_layers, num_input_features, bn_size, growth_rate, drop_rate, efficient=False):
super(_DenseBlock, self).__init__()
for i in range(num_layers):
layer = _DenseLayer(
num_input_features + i * growth_rate,
growth_rate=growth_rate,
bn_size=bn_size,
drop_rate=drop_rate,
efficient=efficient,
)
self.add_module('denselayer%d' % (i + 1), layer)
def forward(self, init_features):
features = [init_features]
for name, layer in self.named_children():
new_features = layer(*features)
features.append(new_features)
return torch.cat(features, 1)
class SE_Block(nn.Module):
def __init__(self, ch_in, reduction=16):
super(SE_Block, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1) # 全局自适应池化
self.fc = nn.Sequential(
nn.Linear(ch_in, ch_in // reduction, bias=False),
nn.ReLU(inplace=True),
nn.Linear(ch_in // reduction, ch_in, bias=False),
nn.Sigmoid()
)
def forward(self, x):
b, c, _, _ = x.size()
y = self.avg_pool(x).view(b, c) # squeeze操作
y = self.fc(y).view(b, c, 1, 1) # FC获取通道注意力权重，是具有全局信息的
return x * y.expand_as(x) # 注意力作用每一个通道上
class DenseNet(nn.Module):
def __init__(self, growth_rate, block_config,num_init_features=24, compression=0.5, bn_size=4, drop_rate=0,
num_classes=10, small_inputs=True, efficient=False):
super(DenseNet, self).__init__()
assert 0 < compression <= 1, 'compression of densenet should be between 0 and 1'
# First convolution
if small_inputs:
self.features = nn.Sequential(OrderedDict([
('conv0', nn.Conv2d(3, num_init_features, kernel_size=3, stride=1, padding=1, bias=False)),
]))
else:
self.features = nn.Sequential(OrderedDict([
('conv0', nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False)),
]))
self.features.add_module('norm0', nn.BatchNorm2d(num_init_features))
self.features.add_module('relu0', nn.ReLU(inplace=True))
self.features.add_module('pool0', nn.MaxPool2d(kernel_size=3, stride=2, padding=1,
ceil_mode=False))
# Each denseblock
num_features = num_init_features
for i, num_layers in enumerate(block_config):
block = _DenseBlock(
num_layers=num_layers,
num_input_features=num_features,
bn_size=bn_size,
growth_rate=growth_rate,
drop_rate=drop_rate,
efficient=efficient,
)
self.features.add_module('denseblock%d' % (i + 1), block)
num_features = num_features + num_layers * growth_rate
if i != len(block_config) - 1:
trans = _Transition(num_input_features=num_features,
num_output_features=int(num_features * compression))
self.features.add_module('transition%d' % (i + 1), trans)
num_features = int(num_features * compression)
#self.features.add_module('SE_Block%d' % (i + 1),SE_Block(num_features, reduction=16))
# Final batch norm
self.features.add_module('norm_final', nn.BatchNorm2d(num_features))
# Linear layer
self.classifier = nn.Linear(num_features, num_classes)
def forward(self, x):
features = self.features(x)
out = F.relu(features, inplace=True)
out = F.adaptive_avg_pool2d(out, (1, 1))
out = torch.flatten(out, 1)
out = self.classifier(out)
return outIn [10]:
# 定义完成，测试一下
x = torch.randn(2, 3, 224, 224)
model = DenseNet(growth_rate=32, block_config=(6,12,24,16), compression=0.5,
num_init_features=64, bn_size=4, drop_rate=0.2,num_classes=4,efficient=True)
out = model(x)
print('out.shape: ', out.shape)
print(out)out.shape: torch.Size([2, 4])
tensor([[-0.0832, -0.0252, 0.1274, 0.2347],
[-0.0983, -0.0233, 0.1142, 0.2350]], grad_fn=<AddmmBackward0>)In [11]:
model.to(device)
# 统计模型参数量以及其他指标
import torchsummary as summary
summary.summary(model, (3, 224, 224))----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 224, 224] 1,728
BatchNorm2d-2 [-1, 64, 224, 224] 128
ReLU-3 [-1, 64, 224, 224] 0
Conv2d-4 [-1, 128, 224, 224] 8,192
BatchNorm2d-5 [-1, 128, 224, 224] 256
ReLU-6 [-1, 128, 224, 224] 0
Conv2d-7 [-1, 32, 224, 224] 36,864
AdaptiveAvgPool2d-8 [-1, 32, 1, 1] 0
Linear-9 [-1, 2] 64
ReLU-10 [-1, 2] 0
Linear-11 [-1, 32] 64
Sigmoid-12 [-1, 32] 0
SE_Block-13 [-1, 32, 224, 224] 0
_DenseLayer-14 [-1, 32, 224, 224] 0
BatchNorm2d-15 [-1, 96, 224, 224] 192
ReLU-16 [-1, 96, 224, 224] 0
Conv2d-17 [-1, 128, 224, 224] 12,288
BatchNorm2d-18 [-1, 128, 224, 224] 256
ReLU-19 [-1, 128, 224, 224] 0
Conv2d-20 [-1, 32, 224, 224] 36,864
AdaptiveAvgPool2d-21 [-1, 32, 1, 1] 0
Linear-22 [-1, 2] 64
ReLU-23 [-1, 2] 0
Linear-24 [-1, 32] 64
Sigmoid-25 [-1, 32] 0
SE_Block-26 [-1, 32, 224, 224] 0
_DenseLayer-27 [-1, 32, 224, 224] 0
BatchNorm2d-28 [-1, 128, 224, 224] 256
ReLU-29 [-1, 128, 224, 224] 0
Conv2d-30 [-1, 128, 224, 224] 16,384
BatchNorm2d-31 [-1, 128, 224, 224] 256
ReLU-32 [-1, 128, 224, 224] 0
Conv2d-33 [-1, 32, 224, 224] 36,864
AdaptiveAvgPool2d-34 [-1, 32, 1, 1] 0
Linear-35 [-1, 2] 64
ReLU-36 [-1, 2] 0
Linear-37 [-1, 32] 64
Sigmoid-38 [-1, 32] 0
SE_Block-39 [-1, 32, 224, 224] 0
_DenseLayer-40 [-1, 32, 224, 224] 0
BatchNorm2d-41 [-1, 160, 224, 224] 320
ReLU-42 [-1, 160, 224, 224] 0
Conv2d-43 [-1, 128, 224, 224] 20,480
BatchNorm2d-44 [-1, 128, 224, 224] 256
ReLU-45 [-1, 128, 224, 224] 0
Conv2d-46 [-1, 32, 224, 224] 36,864
AdaptiveAvgPool2d-47 [-1, 32, 1, 1] 0
Linear-48 [-1, 2] 64
ReLU-49 [-1, 2] 0
Linear-50 [-1, 32] 64
Sigmoid-51 [-1, 32] 0
SE_Block-52 [-1, 32, 224, 224] 0
_DenseLayer-53 [-1, 32, 224, 224] 0
BatchNorm2d-54 [-1, 192, 224, 224] 384
ReLU-55 [-1, 192, 224, 224] 0
Conv2d-56 [-1, 128, 224, 224] 24,576
BatchNorm2d-57 [-1, 128, 224, 224] 256
ReLU-58 [-1, 128, 224, 224] 0
Conv2d-59 [-1, 32, 224, 224] 36,864
AdaptiveAvgPool2d-60 [-1, 32, 1, 1] 0
Linear-61 [-1, 2] 64
ReLU-62 [-1, 2] 0
Linear-63 [-1, 32] 64
Sigmoid-64 [-1, 32] 0
SE_Block-65 [-1, 32, 224, 224] 0
_DenseLayer-66 [-1, 32, 224, 224] 0
BatchNorm2d-67 [-1, 224, 224, 224] 448
ReLU-68 [-1, 224, 224, 224] 0
Conv2d-69 [-1, 128, 224, 224] 28,672
BatchNorm2d-70 [-1, 128, 224, 224] 256
ReLU-71 [-1, 128, 224, 224] 0
Conv2d-72 [-1, 32, 224, 224] 36,864
AdaptiveAvgPool2d-73 [-1, 32, 1, 1] 0
Linear-74 [-1, 2] 64
ReLU-75 [-1, 2] 0
Linear-76 [-1, 32] 64
Sigmoid-77 [-1, 32] 0
SE_Block-78 [-1, 32, 224, 224] 0
_DenseLayer-79 [-1, 32, 224, 224] 0
_DenseBlock-80 [-1, 256, 224, 224] 0
BatchNorm2d-81 [-1, 256, 224, 224] 512
ReLU-82 [-1, 256, 224, 224] 0
Conv2d-83 [-1, 128, 224, 224] 32,768
AvgPool2d-84 [-1, 128, 112, 112] 0
BatchNorm2d-85 [-1, 128, 112, 112] 256
ReLU-86 [-1, 128, 112, 112] 0
Conv2d-87 [-1, 128, 112, 112] 16,384
BatchNorm2d-88 [-1, 128, 112, 112] 256
ReLU-89 [-1, 128, 112, 112] 0
Conv2d-90 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-91 [-1, 32, 1, 1] 0
Linear-92 [-1, 2] 64
ReLU-93 [-1, 2] 0
Linear-94 [-1, 32] 64
Sigmoid-95 [-1, 32] 0
SE_Block-96 [-1, 32, 112, 112] 0
_DenseLayer-97 [-1, 32, 112, 112] 0
BatchNorm2d-98 [-1, 160, 112, 112] 320
ReLU-99 [-1, 160, 112, 112] 0
Conv2d-100 [-1, 128, 112, 112] 20,480
BatchNorm2d-101 [-1, 128, 112, 112] 256
ReLU-102 [-1, 128, 112, 112] 0
Conv2d-103 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-104 [-1, 32, 1, 1] 0
Linear-105 [-1, 2] 64
ReLU-106 [-1, 2] 0
Linear-107 [-1, 32] 64
Sigmoid-108 [-1, 32] 0
SE_Block-109 [-1, 32, 112, 112] 0
_DenseLayer-110 [-1, 32, 112, 112] 0
BatchNorm2d-111 [-1, 192, 112, 112] 384
ReLU-112 [-1, 192, 112, 112] 0
Conv2d-113 [-1, 128, 112, 112] 24,576
BatchNorm2d-114 [-1, 128, 112, 112] 256
ReLU-115 [-1, 128, 112, 112] 0
Conv2d-116 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-117 [-1, 32, 1, 1] 0
Linear-118 [-1, 2] 64
ReLU-119 [-1, 2] 0
Linear-120 [-1, 32] 64
Sigmoid-121 [-1, 32] 0
SE_Block-122 [-1, 32, 112, 112] 0
_DenseLayer-123 [-1, 32, 112, 112] 0
BatchNorm2d-124 [-1, 224, 112, 112] 448
ReLU-125 [-1, 224, 112, 112] 0
Conv2d-126 [-1, 128, 112, 112] 28,672
BatchNorm2d-127 [-1, 128, 112, 112] 256
ReLU-128 [-1, 128, 112, 112] 0
Conv2d-129 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-130 [-1, 32, 1, 1] 0
Linear-131 [-1, 2] 64
ReLU-132 [-1, 2] 0
Linear-133 [-1, 32] 64
Sigmoid-134 [-1, 32] 0
SE_Block-135 [-1, 32, 112, 112] 0
_DenseLayer-136 [-1, 32, 112, 112] 0
BatchNorm2d-137 [-1, 256, 112, 112] 512
ReLU-138 [-1, 256, 112, 112] 0
Conv2d-139 [-1, 128, 112, 112] 32,768
BatchNorm2d-140 [-1, 128, 112, 112] 256
ReLU-141 [-1, 128, 112, 112] 0
Conv2d-142 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-143 [-1, 32, 1, 1] 0
Linear-144 [-1, 2] 64
ReLU-145 [-1, 2] 0
Linear-146 [-1, 32] 64
Sigmoid-147 [-1, 32] 0
SE_Block-148 [-1, 32, 112, 112] 0
_DenseLayer-149 [-1, 32, 112, 112] 0
BatchNorm2d-150 [-1, 288, 112, 112] 576
ReLU-151 [-1, 288, 112, 112] 0
Conv2d-152 [-1, 128, 112, 112] 36,864
BatchNorm2d-153 [-1, 128, 112, 112] 256
ReLU-154 [-1, 128, 112, 112] 0
Conv2d-155 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-156 [-1, 32, 1, 1] 0
Linear-157 [-1, 2] 64
ReLU-158 [-1, 2] 0
Linear-159 [-1, 32] 64
Sigmoid-160 [-1, 32] 0
SE_Block-161 [-1, 32, 112, 112] 0
_DenseLayer-162 [-1, 32, 112, 112] 0
BatchNorm2d-163 [-1, 320, 112, 112] 640
ReLU-164 [-1, 320, 112, 112] 0
Conv2d-165 [-1, 128, 112, 112] 40,960
BatchNorm2d-166 [-1, 128, 112, 112] 256
ReLU-167 [-1, 128, 112, 112] 0
Conv2d-168 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-169 [-1, 32, 1, 1] 0
Linear-170 [-1, 2] 64
ReLU-171 [-1, 2] 0
Linear-172 [-1, 32] 64
Sigmoid-173 [-1, 32] 0
SE_Block-174 [-1, 32, 112, 112] 0
_DenseLayer-175 [-1, 32, 112, 112] 0
BatchNorm2d-176 [-1, 352, 112, 112] 704
ReLU-177 [-1, 352, 112, 112] 0
Conv2d-178 [-1, 128, 112, 112] 45,056
BatchNorm2d-179 [-1, 128, 112, 112] 256
ReLU-180 [-1, 128, 112, 112] 0
Conv2d-181 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-182 [-1, 32, 1, 1] 0
Linear-183 [-1, 2] 64
ReLU-184 [-1, 2] 0
Linear-185 [-1, 32] 64
Sigmoid-186 [-1, 32] 0
SE_Block-187 [-1, 32, 112, 112] 0
_DenseLayer-188 [-1, 32, 112, 112] 0
BatchNorm2d-189 [-1, 384, 112, 112] 768
ReLU-190 [-1, 384, 112, 112] 0
Conv2d-191 [-1, 128, 112, 112] 49,152
BatchNorm2d-192 [-1, 128, 112, 112] 256
ReLU-193 [-1, 128, 112, 112] 0
Conv2d-194 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-195 [-1, 32, 1, 1] 0
Linear-196 [-1, 2] 64
ReLU-197 [-1, 2] 0
Linear-198 [-1, 32] 64
Sigmoid-199 [-1, 32] 0
SE_Block-200 [-1, 32, 112, 112] 0
_DenseLayer-201 [-1, 32, 112, 112] 0
BatchNorm2d-202 [-1, 416, 112, 112] 832
ReLU-203 [-1, 416, 112, 112] 0
Conv2d-204 [-1, 128, 112, 112] 53,248
BatchNorm2d-205 [-1, 128, 112, 112] 256
ReLU-206 [-1, 128, 112, 112] 0
Conv2d-207 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-208 [-1, 32, 1, 1] 0
Linear-209 [-1, 2] 64
ReLU-210 [-1, 2] 0
Linear-211 [-1, 32] 64
Sigmoid-212 [-1, 32] 0
SE_Block-213 [-1, 32, 112, 112] 0
_DenseLayer-214 [-1, 32, 112, 112] 0
BatchNorm2d-215 [-1, 448, 112, 112] 896
ReLU-216 [-1, 448, 112, 112] 0
Conv2d-217 [-1, 128, 112, 112] 57,344
BatchNorm2d-218 [-1, 128, 112, 112] 256
ReLU-219 [-1, 128, 112, 112] 0
Conv2d-220 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-221 [-1, 32, 1, 1] 0
Linear-222 [-1, 2] 64
ReLU-223 [-1, 2] 0
Linear-224 [-1, 32] 64
Sigmoid-225 [-1, 32] 0
SE_Block-226 [-1, 32, 112, 112] 0
_DenseLayer-227 [-1, 32, 112, 112] 0
BatchNorm2d-228 [-1, 480, 112, 112] 960
ReLU-229 [-1, 480, 112, 112] 0
Conv2d-230 [-1, 128, 112, 112] 61,440
BatchNorm2d-231 [-1, 128, 112, 112] 256
ReLU-232 [-1, 128, 112, 112] 0
Conv2d-233 [-1, 32, 112, 112] 36,864
AdaptiveAvgPool2d-234 [-1, 32, 1, 1] 0
Linear-235 [-1, 2] 64
ReLU-236 [-1, 2] 0
Linear-237 [-1, 32] 64
Sigmoid-238 [-1, 32] 0
SE_Block-239 [-1, 32, 112, 112] 0
_DenseLayer-240 [-1, 32, 112, 112] 0
_DenseBlock-241 [-1, 512, 112, 112] 0
BatchNorm2d-242 [-1, 512, 112, 112] 1,024
ReLU-243 [-1, 512, 112, 112] 0
Conv2d-244 [-1, 256, 112, 112] 131,072
AvgPool2d-245 [-1, 256, 56, 56] 0
BatchNorm2d-246 [-1, 256, 56, 56] 512
ReLU-247 [-1, 256, 56, 56] 0
Conv2d-248 [-1, 128, 56, 56] 32,768
BatchNorm2d-249 [-1, 128, 56, 56] 256
ReLU-250 [-1, 128, 56, 56] 0
Conv2d-251 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-252 [-1, 32, 1, 1] 0
Linear-253 [-1, 2] 64
ReLU-254 [-1, 2] 0
Linear-255 [-1, 32] 64
Sigmoid-256 [-1, 32] 0
SE_Block-257 [-1, 32, 56, 56] 0
_DenseLayer-258 [-1, 32, 56, 56] 0
BatchNorm2d-259 [-1, 288, 56, 56] 576
ReLU-260 [-1, 288, 56, 56] 0
Conv2d-261 [-1, 128, 56, 56] 36,864
BatchNorm2d-262 [-1, 128, 56, 56] 256
ReLU-263 [-1, 128, 56, 56] 0
Conv2d-264 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-265 [-1, 32, 1, 1] 0
Linear-266 [-1, 2] 64
ReLU-267 [-1, 2] 0
Linear-268 [-1, 32] 64
Sigmoid-269 [-1, 32] 0
SE_Block-270 [-1, 32, 56, 56] 0
_DenseLayer-271 [-1, 32, 56, 56] 0
BatchNorm2d-272 [-1, 320, 56, 56] 640
ReLU-273 [-1, 320, 56, 56] 0
Conv2d-274 [-1, 128, 56, 56] 40,960
BatchNorm2d-275 [-1, 128, 56, 56] 256
ReLU-276 [-1, 128, 56, 56] 0
Conv2d-277 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-278 [-1, 32, 1, 1] 0
Linear-279 [-1, 2] 64
ReLU-280 [-1, 2] 0
Linear-281 [-1, 32] 64
Sigmoid-282 [-1, 32] 0
SE_Block-283 [-1, 32, 56, 56] 0
_DenseLayer-284 [-1, 32, 56, 56] 0
BatchNorm2d-285 [-1, 352, 56, 56] 704
ReLU-286 [-1, 352, 56, 56] 0
Conv2d-287 [-1, 128, 56, 56] 45,056
BatchNorm2d-288 [-1, 128, 56, 56] 256
ReLU-289 [-1, 128, 56, 56] 0
Conv2d-290 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-291 [-1, 32, 1, 1] 0
Linear-292 [-1, 2] 64
ReLU-293 [-1, 2] 0
Linear-294 [-1, 32] 64
Sigmoid-295 [-1, 32] 0
SE_Block-296 [-1, 32, 56, 56] 0
_DenseLayer-297 [-1, 32, 56, 56] 0
BatchNorm2d-298 [-1, 384, 56, 56] 768
ReLU-299 [-1, 384, 56, 56] 0
Conv2d-300 [-1, 128, 56, 56] 49,152
BatchNorm2d-301 [-1, 128, 56, 56] 256
ReLU-302 [-1, 128, 56, 56] 0
Conv2d-303 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-304 [-1, 32, 1, 1] 0
Linear-305 [-1, 2] 64
ReLU-306 [-1, 2] 0
Linear-307 [-1, 32] 64
Sigmoid-308 [-1, 32] 0
SE_Block-309 [-1, 32, 56, 56] 0
_DenseLayer-310 [-1, 32, 56, 56] 0
BatchNorm2d-311 [-1, 416, 56, 56] 832
ReLU-312 [-1, 416, 56, 56] 0
Conv2d-313 [-1, 128, 56, 56] 53,248
BatchNorm2d-314 [-1, 128, 56, 56] 256
ReLU-315 [-1, 128, 56, 56] 0
Conv2d-316 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-317 [-1, 32, 1, 1] 0
Linear-318 [-1, 2] 64
ReLU-319 [-1, 2] 0
Linear-320 [-1, 32] 64
Sigmoid-321 [-1, 32] 0
SE_Block-322 [-1, 32, 56, 56] 0
_DenseLayer-323 [-1, 32, 56, 56] 0
BatchNorm2d-324 [-1, 448, 56, 56] 896
ReLU-325 [-1, 448, 56, 56] 0
Conv2d-326 [-1, 128, 56, 56] 57,344
BatchNorm2d-327 [-1, 128, 56, 56] 256
ReLU-328 [-1, 128, 56, 56] 0
Conv2d-329 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-330 [-1, 32, 1, 1] 0
Linear-331 [-1, 2] 64
ReLU-332 [-1, 2] 0
Linear-333 [-1, 32] 64
Sigmoid-334 [-1, 32] 0
SE_Block-335 [-1, 32, 56, 56] 0
_DenseLayer-336 [-1, 32, 56, 56] 0
BatchNorm2d-337 [-1, 480, 56, 56] 960
ReLU-338 [-1, 480, 56, 56] 0
Conv2d-339 [-1, 128, 56, 56] 61,440
BatchNorm2d-340 [-1, 128, 56, 56] 256
ReLU-341 [-1, 128, 56, 56] 0
Conv2d-342 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-343 [-1, 32, 1, 1] 0
Linear-344 [-1, 2] 64
ReLU-345 [-1, 2] 0
Linear-346 [-1, 32] 64
Sigmoid-347 [-1, 32] 0
SE_Block-348 [-1, 32, 56, 56] 0
_DenseLayer-349 [-1, 32, 56, 56] 0
BatchNorm2d-350 [-1, 512, 56, 56] 1,024
ReLU-351 [-1, 512, 56, 56] 0
Conv2d-352 [-1, 128, 56, 56] 65,536
BatchNorm2d-353 [-1, 128, 56, 56] 256
ReLU-354 [-1, 128, 56, 56] 0
Conv2d-355 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-356 [-1, 32, 1, 1] 0
Linear-357 [-1, 2] 64
ReLU-358 [-1, 2] 0
Linear-359 [-1, 32] 64
Sigmoid-360 [-1, 32] 0
SE_Block-361 [-1, 32, 56, 56] 0
_DenseLayer-362 [-1, 32, 56, 56] 0
BatchNorm2d-363 [-1, 544, 56, 56] 1,088
ReLU-364 [-1, 544, 56, 56] 0
Conv2d-365 [-1, 128, 56, 56] 69,632
BatchNorm2d-366 [-1, 128, 56, 56] 256
ReLU-367 [-1, 128, 56, 56] 0
Conv2d-368 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-369 [-1, 32, 1, 1] 0
Linear-370 [-1, 2] 64
ReLU-371 [-1, 2] 0
Linear-372 [-1, 32] 64
Sigmoid-373 [-1, 32] 0
SE_Block-374 [-1, 32, 56, 56] 0
_DenseLayer-375 [-1, 32, 56, 56] 0
BatchNorm2d-376 [-1, 576, 56, 56] 1,152
ReLU-377 [-1, 576, 56, 56] 0
Conv2d-378 [-1, 128, 56, 56] 73,728
BatchNorm2d-379 [-1, 128, 56, 56] 256
ReLU-380 [-1, 128, 56, 56] 0
Conv2d-381 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-382 [-1, 32, 1, 1] 0
Linear-383 [-1, 2] 64
ReLU-384 [-1, 2] 0
Linear-385 [-1, 32] 64
Sigmoid-386 [-1, 32] 0
SE_Block-387 [-1, 32, 56, 56] 0
_DenseLayer-388 [-1, 32, 56, 56] 0
BatchNorm2d-389 [-1, 608, 56, 56] 1,216
ReLU-390 [-1, 608, 56, 56] 0
Conv2d-391 [-1, 128, 56, 56] 77,824
BatchNorm2d-392 [-1, 128, 56, 56] 256
ReLU-393 [-1, 128, 56, 56] 0
Conv2d-394 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-395 [-1, 32, 1, 1] 0
Linear-396 [-1, 2] 64
ReLU-397 [-1, 2] 0
Linear-398 [-1, 32] 64
Sigmoid-399 [-1, 32] 0
SE_Block-400 [-1, 32, 56, 56] 0
_DenseLayer-401 [-1, 32, 56, 56] 0
BatchNorm2d-402 [-1, 640, 56, 56] 1,280
ReLU-403 [-1, 640, 56, 56] 0
Conv2d-404 [-1, 128, 56, 56] 81,920
BatchNorm2d-405 [-1, 128, 56, 56] 256
ReLU-406 [-1, 128, 56, 56] 0
Conv2d-407 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-408 [-1, 32, 1, 1] 0
Linear-409 [-1, 2] 64
ReLU-410 [-1, 2] 0
Linear-411 [-1, 32] 64
Sigmoid-412 [-1, 32] 0
SE_Block-413 [-1, 32, 56, 56] 0
_DenseLayer-414 [-1, 32, 56, 56] 0
BatchNorm2d-415 [-1, 672, 56, 56] 1,344
ReLU-416 [-1, 672, 56, 56] 0
Conv2d-417 [-1, 128, 56, 56] 86,016
BatchNorm2d-418 [-1, 128, 56, 56] 256
ReLU-419 [-1, 128, 56, 56] 0
Conv2d-420 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-421 [-1, 32, 1, 1] 0
Linear-422 [-1, 2] 64
ReLU-423 [-1, 2] 0
Linear-424 [-1, 32] 64
Sigmoid-425 [-1, 32] 0
SE_Block-426 [-1, 32, 56, 56] 0
_DenseLayer-427 [-1, 32, 56, 56] 0
BatchNorm2d-428 [-1, 704, 56, 56] 1,408
ReLU-429 [-1, 704, 56, 56] 0
Conv2d-430 [-1, 128, 56, 56] 90,112
BatchNorm2d-431 [-1, 128, 56, 56] 256
ReLU-432 [-1, 128, 56, 56] 0
Conv2d-433 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-434 [-1, 32, 1, 1] 0
Linear-435 [-1, 2] 64
ReLU-436 [-1, 2] 0
Linear-437 [-1, 32] 64
Sigmoid-438 [-1, 32] 0
SE_Block-439 [-1, 32, 56, 56] 0
_DenseLayer-440 [-1, 32, 56, 56] 0
BatchNorm2d-441 [-1, 736, 56, 56] 1,472
ReLU-442 [-1, 736, 56, 56] 0
Conv2d-443 [-1, 128, 56, 56] 94,208
BatchNorm2d-444 [-1, 128, 56, 56] 256
ReLU-445 [-1, 128, 56, 56] 0
Conv2d-446 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-447 [-1, 32, 1, 1] 0
Linear-448 [-1, 2] 64
ReLU-449 [-1, 2] 0
Linear-450 [-1, 32] 64
Sigmoid-451 [-1, 32] 0
SE_Block-452 [-1, 32, 56, 56] 0
_DenseLayer-453 [-1, 32, 56, 56] 0
BatchNorm2d-454 [-1, 768, 56, 56] 1,536
ReLU-455 [-1, 768, 56, 56] 0
Conv2d-456 [-1, 128, 56, 56] 98,304
BatchNorm2d-457 [-1, 128, 56, 56] 256
ReLU-458 [-1, 128, 56, 56] 0
Conv2d-459 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-460 [-1, 32, 1, 1] 0
Linear-461 [-1, 2] 64
ReLU-462 [-1, 2] 0
Linear-463 [-1, 32] 64
Sigmoid-464 [-1, 32] 0
SE_Block-465 [-1, 32, 56, 56] 0
_DenseLayer-466 [-1, 32, 56, 56] 0
BatchNorm2d-467 [-1, 800, 56, 56] 1,600
ReLU-468 [-1, 800, 56, 56] 0
Conv2d-469 [-1, 128, 56, 56] 102,400
BatchNorm2d-470 [-1, 128, 56, 56] 256
ReLU-471 [-1, 128, 56, 56] 0
Conv2d-472 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-473 [-1, 32, 1, 1] 0
Linear-474 [-1, 2] 64
ReLU-475 [-1, 2] 0
Linear-476 [-1, 32] 64
Sigmoid-477 [-1, 32] 0
SE_Block-478 [-1, 32, 56, 56] 0
_DenseLayer-479 [-1, 32, 56, 56] 0
BatchNorm2d-480 [-1, 832, 56, 56] 1,664
ReLU-481 [-1, 832, 56, 56] 0
Conv2d-482 [-1, 128, 56, 56] 106,496
BatchNorm2d-483 [-1, 128, 56, 56] 256
ReLU-484 [-1, 128, 56, 56] 0
Conv2d-485 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-486 [-1, 32, 1, 1] 0
Linear-487 [-1, 2] 64
ReLU-488 [-1, 2] 0
Linear-489 [-1, 32] 64
Sigmoid-490 [-1, 32] 0
SE_Block-491 [-1, 32, 56, 56] 0
_DenseLayer-492 [-1, 32, 56, 56] 0
BatchNorm2d-493 [-1, 864, 56, 56] 1,728
ReLU-494 [-1, 864, 56, 56] 0
Conv2d-495 [-1, 128, 56, 56] 110,592
BatchNorm2d-496 [-1, 128, 56, 56] 256
ReLU-497 [-1, 128, 56, 56] 0
Conv2d-498 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-499 [-1, 32, 1, 1] 0
Linear-500 [-1, 2] 64
ReLU-501 [-1, 2] 0
Linear-502 [-1, 32] 64
Sigmoid-503 [-1, 32] 0
SE_Block-504 [-1, 32, 56, 56] 0
_DenseLayer-505 [-1, 32, 56, 56] 0
BatchNorm2d-506 [-1, 896, 56, 56] 1,792
ReLU-507 [-1, 896, 56, 56] 0
Conv2d-508 [-1, 128, 56, 56] 114,688
BatchNorm2d-509 [-1, 128, 56, 56] 256
ReLU-510 [-1, 128, 56, 56] 0
Conv2d-511 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-512 [-1, 32, 1, 1] 0
Linear-513 [-1, 2] 64
ReLU-514 [-1, 2] 0
Linear-515 [-1, 32] 64
Sigmoid-516 [-1, 32] 0
SE_Block-517 [-1, 32, 56, 56] 0
_DenseLayer-518 [-1, 32, 56, 56] 0
BatchNorm2d-519 [-1, 928, 56, 56] 1,856
ReLU-520 [-1, 928, 56, 56] 0
Conv2d-521 [-1, 128, 56, 56] 118,784
BatchNorm2d-522 [-1, 128, 56, 56] 256
ReLU-523 [-1, 128, 56, 56] 0
Conv2d-524 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-525 [-1, 32, 1, 1] 0
Linear-526 [-1, 2] 64
ReLU-527 [-1, 2] 0
Linear-528 [-1, 32] 64
Sigmoid-529 [-1, 32] 0
SE_Block-530 [-1, 32, 56, 56] 0
_DenseLayer-531 [-1, 32, 56, 56] 0
BatchNorm2d-532 [-1, 960, 56, 56] 1,920
ReLU-533 [-1, 960, 56, 56] 0
Conv2d-534 [-1, 128, 56, 56] 122,880
BatchNorm2d-535 [-1, 128, 56, 56] 256
ReLU-536 [-1, 128, 56, 56] 0
Conv2d-537 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-538 [-1, 32, 1, 1] 0
Linear-539 [-1, 2] 64
ReLU-540 [-1, 2] 0
Linear-541 [-1, 32] 64
Sigmoid-542 [-1, 32] 0
SE_Block-543 [-1, 32, 56, 56] 0
_DenseLayer-544 [-1, 32, 56, 56] 0
BatchNorm2d-545 [-1, 992, 56, 56] 1,984
ReLU-546 [-1, 992, 56, 56] 0
Conv2d-547 [-1, 128, 56, 56] 126,976
BatchNorm2d-548 [-1, 128, 56, 56] 256
ReLU-549 [-1, 128, 56, 56] 0
Conv2d-550 [-1, 32, 56, 56] 36,864
AdaptiveAvgPool2d-551 [-1, 32, 1, 1] 0
Linear-552 [-1, 2] 64
ReLU-553 [-1, 2] 0
Linear-554 [-1, 32] 64
Sigmoid-555 [-1, 32] 0
SE_Block-556 [-1, 32, 56, 56] 0
_DenseLayer-557 [-1, 32, 56, 56] 0
_DenseBlock-558 [-1, 1024, 56, 56] 0
BatchNorm2d-559 [-1, 1024, 56, 56] 2,048
ReLU-560 [-1, 1024, 56, 56] 0
Conv2d-561 [-1, 512, 56, 56] 524,288
AvgPool2d-562 [-1, 512, 28, 28] 0
BatchNorm2d-563 [-1, 512, 28, 28] 1,024
ReLU-564 [-1, 512, 28, 28] 0
Conv2d-565 [-1, 128, 28, 28] 65,536
BatchNorm2d-566 [-1, 128, 28, 28] 256
ReLU-567 [-1, 128, 28, 28] 0
Conv2d-568 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-569 [-1, 32, 1, 1] 0
Linear-570 [-1, 2] 64
ReLU-571 [-1, 2] 0
Linear-572 [-1, 32] 64
Sigmoid-573 [-1, 32] 0
SE_Block-574 [-1, 32, 28, 28] 0
_DenseLayer-575 [-1, 32, 28, 28] 0
BatchNorm2d-576 [-1, 544, 28, 28] 1,088
ReLU-577 [-1, 544, 28, 28] 0
Conv2d-578 [-1, 128, 28, 28] 69,632
BatchNorm2d-579 [-1, 128, 28, 28] 256
ReLU-580 [-1, 128, 28, 28] 0
Conv2d-581 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-582 [-1, 32, 1, 1] 0
Linear-583 [-1, 2] 64
ReLU-584 [-1, 2] 0
Linear-585 [-1, 32] 64
Sigmoid-586 [-1, 32] 0
SE_Block-587 [-1, 32, 28, 28] 0
_DenseLayer-588 [-1, 32, 28, 28] 0
BatchNorm2d-589 [-1, 576, 28, 28] 1,152
ReLU-590 [-1, 576, 28, 28] 0
Conv2d-591 [-1, 128, 28, 28] 73,728
BatchNorm2d-592 [-1, 128, 28, 28] 256
ReLU-593 [-1, 128, 28, 28] 0
Conv2d-594 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-595 [-1, 32, 1, 1] 0
Linear-596 [-1, 2] 64
ReLU-597 [-1, 2] 0
Linear-598 [-1, 32] 64
Sigmoid-599 [-1, 32] 0
SE_Block-600 [-1, 32, 28, 28] 0
_DenseLayer-601 [-1, 32, 28, 28] 0
BatchNorm2d-602 [-1, 608, 28, 28] 1,216
ReLU-603 [-1, 608, 28, 28] 0
Conv2d-604 [-1, 128, 28, 28] 77,824
BatchNorm2d-605 [-1, 128, 28, 28] 256
ReLU-606 [-1, 128, 28, 28] 0
Conv2d-607 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-608 [-1, 32, 1, 1] 0
Linear-609 [-1, 2] 64
ReLU-610 [-1, 2] 0
Linear-611 [-1, 32] 64
Sigmoid-612 [-1, 32] 0
SE_Block-613 [-1, 32, 28, 28] 0
_DenseLayer-614 [-1, 32, 28, 28] 0
BatchNorm2d-615 [-1, 640, 28, 28] 1,280
ReLU-616 [-1, 640, 28, 28] 0
Conv2d-617 [-1, 128, 28, 28] 81,920
BatchNorm2d-618 [-1, 128, 28, 28] 256
ReLU-619 [-1, 128, 28, 28] 0
Conv2d-620 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-621 [-1, 32, 1, 1] 0
Linear-622 [-1, 2] 64
ReLU-623 [-1, 2] 0
Linear-624 [-1, 32] 64
Sigmoid-625 [-1, 32] 0
SE_Block-626 [-1, 32, 28, 28] 0
_DenseLayer-627 [-1, 32, 28, 28] 0
BatchNorm2d-628 [-1, 672, 28, 28] 1,344
ReLU-629 [-1, 672, 28, 28] 0
Conv2d-630 [-1, 128, 28, 28] 86,016
BatchNorm2d-631 [-1, 128, 28, 28] 256
ReLU-632 [-1, 128, 28, 28] 0
Conv2d-633 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-634 [-1, 32, 1, 1] 0
Linear-635 [-1, 2] 64
ReLU-636 [-1, 2] 0
Linear-637 [-1, 32] 64
Sigmoid-638 [-1, 32] 0
SE_Block-639 [-1, 32, 28, 28] 0
_DenseLayer-640 [-1, 32, 28, 28] 0
BatchNorm2d-641 [-1, 704, 28, 28] 1,408
ReLU-642 [-1, 704, 28, 28] 0
Conv2d-643 [-1, 128, 28, 28] 90,112
BatchNorm2d-644 [-1, 128, 28, 28] 256
ReLU-645 [-1, 128, 28, 28] 0
Conv2d-646 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-647 [-1, 32, 1, 1] 0
Linear-648 [-1, 2] 64
ReLU-649 [-1, 2] 0
Linear-650 [-1, 32] 64
Sigmoid-651 [-1, 32] 0
SE_Block-652 [-1, 32, 28, 28] 0
_DenseLayer-653 [-1, 32, 28, 28] 0
BatchNorm2d-654 [-1, 736, 28, 28] 1,472
ReLU-655 [-1, 736, 28, 28] 0
Conv2d-656 [-1, 128, 28, 28] 94,208
BatchNorm2d-657 [-1, 128, 28, 28] 256
ReLU-658 [-1, 128, 28, 28] 0
Conv2d-659 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-660 [-1, 32, 1, 1] 0
Linear-661 [-1, 2] 64
ReLU-662 [-1, 2] 0
Linear-663 [-1, 32] 64
Sigmoid-664 [-1, 32] 0
SE_Block-665 [-1, 32, 28, 28] 0
_DenseLayer-666 [-1, 32, 28, 28] 0
BatchNorm2d-667 [-1, 768, 28, 28] 1,536
ReLU-668 [-1, 768, 28, 28] 0
Conv2d-669 [-1, 128, 28, 28] 98,304
BatchNorm2d-670 [-1, 128, 28, 28] 256
ReLU-671 [-1, 128, 28, 28] 0
Conv2d-672 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-673 [-1, 32, 1, 1] 0
Linear-674 [-1, 2] 64
ReLU-675 [-1, 2] 0
Linear-676 [-1, 32] 64
Sigmoid-677 [-1, 32] 0
SE_Block-678 [-1, 32, 28, 28] 0
_DenseLayer-679 [-1, 32, 28, 28] 0
BatchNorm2d-680 [-1, 800, 28, 28] 1,600
ReLU-681 [-1, 800, 28, 28] 0
Conv2d-682 [-1, 128, 28, 28] 102,400
BatchNorm2d-683 [-1, 128, 28, 28] 256
ReLU-684 [-1, 128, 28, 28] 0
Conv2d-685 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-686 [-1, 32, 1, 1] 0
Linear-687 [-1, 2] 64
ReLU-688 [-1, 2] 0
Linear-689 [-1, 32] 64
Sigmoid-690 [-1, 32] 0
SE_Block-691 [-1, 32, 28, 28] 0
_DenseLayer-692 [-1, 32, 28, 28] 0
BatchNorm2d-693 [-1, 832, 28, 28] 1,664
ReLU-694 [-1, 832, 28, 28] 0
Conv2d-695 [-1, 128, 28, 28] 106,496
BatchNorm2d-696 [-1, 128, 28, 28] 256
ReLU-697 [-1, 128, 28, 28] 0
Conv2d-698 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-699 [-1, 32, 1, 1] 0
Linear-700 [-1, 2] 64
ReLU-701 [-1, 2] 0
Linear-702 [-1, 32] 64
Sigmoid-703 [-1, 32] 0
SE_Block-704 [-1, 32, 28, 28] 0
_DenseLayer-705 [-1, 32, 28, 28] 0
BatchNorm2d-706 [-1, 864, 28, 28] 1,728
ReLU-707 [-1, 864, 28, 28] 0
Conv2d-708 [-1, 128, 28, 28] 110,592
BatchNorm2d-709 [-1, 128, 28, 28] 256
ReLU-710 [-1, 128, 28, 28] 0
Conv2d-711 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-712 [-1, 32, 1, 1] 0
Linear-713 [-1, 2] 64
ReLU-714 [-1, 2] 0
Linear-715 [-1, 32] 64
Sigmoid-716 [-1, 32] 0
SE_Block-717 [-1, 32, 28, 28] 0
_DenseLayer-718 [-1, 32, 28, 28] 0
BatchNorm2d-719 [-1, 896, 28, 28] 1,792
ReLU-720 [-1, 896, 28, 28] 0
Conv2d-721 [-1, 128, 28, 28] 114,688
BatchNorm2d-722 [-1, 128, 28, 28] 256
ReLU-723 [-1, 128, 28, 28] 0
Conv2d-724 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-725 [-1, 32, 1, 1] 0
Linear-726 [-1, 2] 64
ReLU-727 [-1, 2] 0
Linear-728 [-1, 32] 64
Sigmoid-729 [-1, 32] 0
SE_Block-730 [-1, 32, 28, 28] 0
_DenseLayer-731 [-1, 32, 28, 28] 0
BatchNorm2d-732 [-1, 928, 28, 28] 1,856
ReLU-733 [-1, 928, 28, 28] 0
Conv2d-734 [-1, 128, 28, 28] 118,784
BatchNorm2d-735 [-1, 128, 28, 28] 256
ReLU-736 [-1, 128, 28, 28] 0
Conv2d-737 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-738 [-1, 32, 1, 1] 0
Linear-739 [-1, 2] 64
ReLU-740 [-1, 2] 0
Linear-741 [-1, 32] 64
Sigmoid-742 [-1, 32] 0
SE_Block-743 [-1, 32, 28, 28] 0
_DenseLayer-744 [-1, 32, 28, 28] 0
BatchNorm2d-745 [-1, 960, 28, 28] 1,920
ReLU-746 [-1, 960, 28, 28] 0
Conv2d-747 [-1, 128, 28, 28] 122,880
BatchNorm2d-748 [-1, 128, 28, 28] 256
ReLU-749 [-1, 128, 28, 28] 0
Conv2d-750 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-751 [-1, 32, 1, 1] 0
Linear-752 [-1, 2] 64
ReLU-753 [-1, 2] 0
Linear-754 [-1, 32] 64
Sigmoid-755 [-1, 32] 0
SE_Block-756 [-1, 32, 28, 28] 0
_DenseLayer-757 [-1, 32, 28, 28] 0
BatchNorm2d-758 [-1, 992, 28, 28] 1,984
ReLU-759 [-1, 992, 28, 28] 0
Conv2d-760 [-1, 128, 28, 28] 126,976
BatchNorm2d-761 [-1, 128, 28, 28] 256
ReLU-762 [-1, 128, 28, 28] 0
Conv2d-763 [-1, 32, 28, 28] 36,864
AdaptiveAvgPool2d-764 [-1, 32, 1, 1] 0
Linear-765 [-1, 2] 64
ReLU-766 [-1, 2] 0
Linear-767 [-1, 32] 64
Sigmoid-768 [-1, 32] 0
SE_Block-769 [-1, 32, 28, 28] 0
_DenseLayer-770 [-1, 32, 28, 28] 0
_DenseBlock-771 [-1, 1024, 28, 28] 0
BatchNorm2d-772 [-1, 1024, 28, 28] 2,048
Linear-773 [-1, 4] 4,100
================================================================
Total params: 6,957,572
Trainable params: 6,957,572
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 4854.11
Params size (MB): 26.54
Estimated Total Size (MB): 4881.22
----------------------------------------------------------------In [12]:
# 训练循环
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset) # 训练集的大小
num_batches = len(dataloader) # 批次数目, (size/batch_size，向上取整)
train_loss, train_acc = 0, 0 # 初始化训练损失和正确率
for X, y in dataloader: # 获取图片及其标签
X, y = X.to(device), y.to(device)
# 计算预测误差
pred = model(X) # 网络输出
loss = loss_fn(pred, y) # 计算网络输出和真实值之间的差距，targets为真实值，计算二者差值即为损失
# 反向传播
optimizer.zero_grad() # grad属性归零
loss.backward() # 反向传播
optimizer.step() # 每一步自动更新
# 记录acc与loss
train_acc += (pred.argmax(1) == y).type(torch.float).sum().item()
train_loss += loss.item()
train_acc /= size
train_loss /= num_batches
return train_acc, train_lossIn [13]:
def test (dataloader, model, loss_fn):
size = len(dataloader.dataset) # 测试集的大小
num_batches = len(dataloader) # 批次数目, (size/batch_size，向上取整)
test_loss, test_acc = 0, 0
# 当不进行训练时，停止梯度更新，节省计算内存消耗
with torch.no_grad():
for imgs, target in dataloader:
imgs, target = imgs.to(device), target.to(device)
# 计算loss
target_pred = model(imgs)
loss = loss_fn(target_pred, target)
test_loss += loss.item()
test_acc += (target_pred.argmax(1) == target).type(torch.float).sum().item()
test_acc /= size
test_loss /= num_batches
return test_acc, test_lossIn [14]:
learn_rate = 1e-4 # 初始学习率
optimizer = torch.optim.Adam(model.parameters(), lr=learn_rate)In [15]:
import copy
optimizer = torch.optim.Adam(model.parameters(), lr= 1e-4)
loss_fn = nn.CrossEntropyLoss() # 创建损失函数
epochs = 10
train_loss = []
train_acc = []
test_loss = []
test_acc = []
best_acc = 0 # 设置一个最佳准确率，作为最佳模型的判别指标
for epoch in range(epochs):
model.train()
epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)
model.eval()
epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
# 保存最佳模型到 best_model
if epoch_test_acc > best_acc:
best_acc = epoch_test_acc
best_model = copy.deepcopy(model)
train_acc.append(epoch_train_acc)
train_loss.append(epoch_train_loss)
test_acc.append(epoch_test_acc)
test_loss.append(epoch_test_loss)
# 获取当前的学习率
lr = optimizer.state_dict()['param_groups'][0]['lr']
template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E}')
print(template.format(epoch+1, epoch_train_acc*100, epoch_train_loss,
epoch_test_acc*100, epoch_test_loss, lr))
# 保存最佳模型到文件中
# PATH = './best_model.pth' # 保存的参数文件名
# torch.save(model.state_dict(), PATH)
print('Done')Epoch: 1, Train_acc:86.5%, Train_loss:0.403, Test_acc:75.4%, Test_loss:0.699, Lr:1.00E-04
Epoch: 2, Train_acc:89.9%, Train_loss:0.311, Test_acc:95.8%, Test_loss:0.115, Lr:1.00E-04
Epoch: 3, Train_acc:92.3%, Train_loss:0.216, Test_acc:96.7%, Test_loss:0.084, Lr:1.00E-04
Epoch: 4, Train_acc:93.9%, Train_loss:0.189, Test_acc:96.7%, Test_loss:0.136, Lr:1.00E-04
Epoch: 5, Train_acc:93.9%, Train_loss:0.181, Test_acc:98.8%, Test_loss:0.060, Lr:1.00E-04
Epoch: 6, Train_acc:96.9%, Train_loss:0.115, Test_acc:94.2%, Test_loss:0.129, Lr:1.00E-04
Epoch: 7, Train_acc:98.2%, Train_loss:0.080, Test_acc:96.7%, Test_loss:0.094, Lr:1.00E-04
Epoch: 8, Train_acc:96.1%, Train_loss:0.118, Test_acc:97.5%, Test_loss:0.115, Lr:1.00E-04
Epoch: 9, Train_acc:96.6%, Train_loss:0.112, Test_acc:97.9%, Test_loss:0.101, Lr:1.00E-04
Epoch:10, Train_acc:97.8%, Train_loss:0.078, Test_acc:90.8%, Test_loss:0.218, Lr:1.00E-04
DoneIn [ ]:
浙公网安备 33010602011771号