自己动手搭建模型：图像分类篇
本文介绍用PaddlePaddle搭建AlexNet、VGG、ResNet、DenseNet等深度学习模型的过程。先处理数据集，取两类图片划分训练、验证集，定义数据集类并预处理。接着分别构建各模型，展示结构与参数，最后训练验证，其中前三者预测准确，DenseNet略有偏差。
项目简介
对于我们初学者来说，很多时候只关注了模型的调用，对于paddle封装好的完整模型套件，只需要train,predict就可以完成一个深度学习任务，但是如果想进一步深入学习，则需要对模型的结构有所了解，试着自己用API去组建模型是一个很好的方案，因为在搭建的过程中会遇到很多结构图上看起来很简单，但是实现的时候会有点小迷糊的细节，这些细节会让你去更深的思考这些模型，同时在加深对模型理解的同时，还能增强编程能力。
数据集介绍
数据集取自第三届中国AI+创新创业大赛：半监督学习目标定位竞赛，原数据集取自IMAGENET，共100类，各类别分别100图片，为了便于训练和验证，只取其中2类作为数据集。
导入相关库
In [1]
import paddleimport paddle.nn as nnimport numpy as npfrom paddle.io import Dataset,DataLoaderfrom tqdm import tqdmimport osimport randomimport cv2from paddle.vision import transforms as Timport paddle.nn.functional as Fimport pandas as pdimport math
进行相关配置
In [2]
lr_base=1e-3epochs=10batch_size=4image_size=(256,256)
生成数据集文件和数据集
这里的数据集采自IMAGENET，因为只是为了验证模型的正确性，所以数据集比较小，两个类别各100张图片，分别置于image_class1,image_class2文件夹中
In [49]
!unzip -qo dataset.zip
In [3]
dataset=[]with open("dataset.txt","a") as f: for img in os.listdir("image_class1"):
dataset.append("image_class1/"+img+" 0\n")
f.write("image_class1/"+img+" 0\n") for img in os.listdir("image_class2"):
dataset.append("image_class2/"+img+" 1\n")
f.write("image_class2/"+img+" 1\n")
random.shuffle(dataset)
val=0.2offset=int(len(dataset)*val)with open("train_list.txt","a")as f: for img in dataset[:-offset]:
f.write(img)with open("val_list.txt","a")as f: for img in dataset[-offset:]:
f.write(img)
这里提示一下一定要注意数据的归一化预处理，否则模型通常不收敛
In [4]
#datasetn for alexnet c=3class DatasetN(Dataset):
def __init__(self,data="dataset.txt",transforms=None):
super().__init__()
self.dataset=open(data).readlines()
self.dataset=[d.strip() for d in self.dataset]
self.transforms=transforms
def __getitem__(self,ind):
data=self.dataset[ind]
img,label=data.split(" ")
img=cv2.imread(img,1)
img=cv2.resize(img,(224,224)).astype("float32").transpose((2,0,1))
img=img.reshape((-1,224,224)) if img is not None:
img=img#self.transforms(img)
img=(img-127.5)/255.0
label=int(label) return img,label
def __len__(self):
return len(self.dataset)
train_set=DatasetN("train_list.txt",T.Compose([T.Normalize(data_format="CHW")]))
val_set=DatasetN("val_list.txt",T.Compose([T.Normalize(data_format="CHW")]))
train_loader=DataLoader(train_set,batch_size=batch_size,shuffle=True)
val_loader=DataLoader(val_set,batch_size=batch_size)
构建模型
AlexNet
图像分类模型中第一个被广泛注意到的使用卷积的深度学习模型，也是这一波深度学习浪潮的引领者，在2012年的IMAGENET中以远超第二名的成绩夺冠，自此深度学习再次走入人们的视野。这个模型在论文《ImageNet Classification with Deep Convolutional Neural Networks》中被提出。 模型结构如图所示 结构图分成了上下两层，是因为原作者在两块GPU中进行的模型训练，最后进行合并，在理解结构图时只需要看作同一层即可。
根据结构图，可以看出输入是一个3通道的图像，第一层是大小为11的卷积操作，接着是5* 5卷积和MaxPooling，以及连续的3* 3卷积和MaxPooling，最后将特征图接入全连接层，所以要先Flatten，最后一个全连接层的维度是所分类别的数量，在原文中是1000，需要根据自己的数据集进行修改。通常卷积和全连接层后面都会加上ReLU作为激活函数。
In [14]
#AlexNetclass AlexNet(nn.Layer):
def __init__(self):
super().__init__()
self.layers=nn.LayerList()
self.layers.append(nn.Sequential(nn.Conv2D(3,96,kernel_size=11,stride=4),nn.ReLU()))
self.layers.append(nn.Sequential(nn.Conv2D(96,256,kernel_size=5,stride=1,padding=2),nn.ReLU(),nn.MaxPool2D(kernel_size=3,stride=2)))
self.layers.append(nn.Sequential(nn.Conv2D(256,384,kernel_size=3,stride=1,padding=1),nn.ReLU(),nn.MaxPool2D(kernel_size=3,stride=2,padding=0)))
self.layers.append(nn.Sequential(nn.Conv2D(384,384,kernel_size=3,stride=1,padding=1),nn.ReLU()))
self.layers.append(nn.Sequential(nn.Conv2D(384,256,kernel_size=3,stride=1,padding=1),nn.ReLU(),nn.MaxPool2D(kernel_size=3,stride=2)))
self.layers.append(nn.Sequential(nn.Flatten(),nn.Linear(6400,4096),nn.ReLU()))
self.layers.append(nn.Sequential(nn.Linear(4096,4096),nn.ReLU()))
self.layers.append(nn.Sequential(nn.Linear(4096,2)))
def forward(self,x):
for layer in self.layers:
y=layer(x)
x=y return y
network=AlexNet()
paddle.summary(network,(16,3,224,224))#使用paddle API验证结构正确性
---------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
===========================================================================
Conv2D-46 [[16, 3, 224, 224]] [16, 96, 54, 54] 34,944
ReLU-64 [[16, 96, 54, 54]] [16, 96, 54, 54] 0
Conv2D-47 [[16, 96, 54, 54]] [16, 256, 54, 54] 614,656
ReLU-65 [[16, 256, 54, 54]] [16, 256, 54, 54] 0
MaxPool2D-28 [[16, 256, 54, 54]] [16, 256, 26, 26] 0
Conv2D-48 [[16, 256, 26, 26]] [16, 384, 26, 26] 885,120
ReLU-66 [[16, 384, 26, 26]] [16, 384, 26, 26] 0
MaxPool2D-29 [[16, 384, 26, 26]] [16, 384, 12, 12] 0
Conv2D-49 [[16, 384, 12, 12]] [16, 384, 12, 12] 1,327,488
ReLU-67 [[16, 384, 12, 12]] [16, 384, 12, 12] 0
Conv2D-50 [[16, 384, 12, 12]] [16, 256, 12, 12] 884,992
ReLU-68 [[16, 256, 12, 12]] [16, 256, 12, 12] 0
MaxPool2D-30 [[16, 256, 12, 12]] [16, 256, 5, 5] 0
Flatten-17 [[16, 256, 5, 5]] [16, 6400] 0
Linear-28 [[16, 6400]] [16, 4096] 26,218,496
ReLU-69 [[16, 4096]] [16, 4096] 0
Linear-29 [[16, 4096]] [16, 4096] 16,781,312
ReLU-70 [[16, 4096]] [16, 4096] 0
Linear-30 [[16, 4096]] [16, 2] 8,194
===========================================================================
Total params: 46,755,202
Trainable params: 46,755,202
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 9.19
Forward/backward pass size (MB): 367.91
Params size (MB): 178.36
Estimated Total Size (MB): 555.45
---------------------------------------------------------------------------
{'total_params': 46755202, 'trainable_params': 46755202}
在模型训练和验证部分，验证结果为
Predict begin...step 10/10 [==============================] - 5ms/step Predict samples: 10[1 1 0 1 0 0 0 1 0 1]
正确答案为1 1 0 1 0 0 0 1 0 1
VGG Net
VGG网络是继AlexNet之后的另一个卷积网络，主要工作是使用了3* 3的卷积以及更深的网络，其基本结构还是Conv+Pooling，在论文 Very Deep Convolutional Networks for Large-Scale Image Recognition由Oxford Visual Geometry Group提出，常用的结构有VGG11、VGG13、VGG16、VGG19，是2014年ILSVRC竞赛的第二名，第一名是GoogLeNet，但是在很多的迁移学习中，VGG表现更优。模型结构图如下
In [15]
#VGG Netclass VGG(nn.Layer):
def __init__(self,features,num_classes):
super().__init__()
self.features=features
self.cls=nn.Sequential(nn.Dropout(0.5),nn.Linear(7*7*512,4096),nn.ReLU(),nn.Linear(4096,4096),nn.ReLU(),nn.Linear(4096,num_classes))
def forward(self,x):
x=self.features(x)
x=self.cls(x) return xdef make_features(cfg):
layers=[]
in_c=3
for v in cfg: if v=="P":
layers.append(nn.MaxPool2D(2)) else:
layers.append(nn.Conv2D(in_c,v,kernel_size=3,stride=1,padding=1))
layers.append(nn.ReLU())
in_c=v
layers.append(nn.Flatten()) return nn.Sequential(*layers)
cfg={ 'VGG11': [64, 'P', 128, 'P', 256, 256, 'P', 512, 512, 'P', 512, 512, 'P'], 'VGG13': [64, 64, 'P', 128, 128, 'P', 256, 256, 'P', 512, 512, 'P', 512, 512, 'P'], 'VGG16': [64, 64, 'P', 128, 128, 'P', 256, 256, 256, 'P', 512, 512, 512, 'P', 512, 512, 512, 'P'], "VGG19": [64, 64, 'P', 128, 128, 'P', 256, 256, 256, 256, 'P', 512, 512, 512, 512, 'P', 512, 512, 512, 512, 'P']
}
features=make_features(cfg["VGG11"])
network=VGG(features,2)
paddle.summary(network,(16,3,224,224))#使用paddle API验证结构正确性
----------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
============================================================================
Conv2D-51 [[16, 3, 224, 224]] [16, 64, 224, 224] 1,792
ReLU-71 [[16, 64, 224, 224]] [16, 64, 224, 224] 0
MaxPool2D-31 [[16, 64, 224, 224]] [16, 64, 112, 112] 0
Conv2D-52 [[16, 64, 112, 112]] [16, 128, 112, 112] 73,856
ReLU-72 [[16, 128, 112, 112]] [16, 128, 112, 112] 0
MaxPool2D-32 [[16, 128, 112, 112]] [16, 128, 56, 56] 0
Conv2D-53 [[16, 128, 56, 56]] [16, 256, 56, 56] 295,168
ReLU-73 [[16, 256, 56, 56]] [16, 256, 56, 56] 0
Conv2D-54 [[16, 256, 56, 56]] [16, 256, 56, 56] 590,080
ReLU-74 [[16, 256, 56, 56]] [16, 256, 56, 56] 0
MaxPool2D-33 [[16, 256, 56, 56]] [16, 256, 28, 28] 0
Conv2D-55 [[16, 256, 28, 28]] [16, 512, 28, 28] 1,180,160
ReLU-75 [[16, 512, 28, 28]] [16, 512, 28, 28] 0
Conv2D-56 [[16, 512, 28, 28]] [16, 512, 28, 28] 2,359,808
ReLU-76 [[16, 512, 28, 28]] [16, 512, 28, 28] 0
MaxPool2D-34 [[16, 512, 28, 28]] [16, 512, 14, 14] 0
Conv2D-57 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,359,808
ReLU-77 [[16, 512, 14, 14]] [16, 512, 14, 14] 0
Conv2D-58 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,359,808
ReLU-78 [[16, 512, 14, 14]] [16, 512, 14, 14] 0
MaxPool2D-35 [[16, 512, 14, 14]] [16, 512, 7, 7] 0
Flatten-19 [[16, 512, 7, 7]] [16, 25088] 0
Dropout-1 [[16, 25088]] [16, 25088] 0
Linear-31 [[16, 25088]] [16, 4096] 102,764,544
ReLU-79 [[16, 4096]] [16, 4096] 0
Linear-32 [[16, 4096]] [16, 4096] 16,781,312
ReLU-80 [[16, 4096]] [16, 4096] 0
Linear-33 [[16, 4096]] [16, 2] 8,194
============================================================================
Total params: 128,774,530
Trainable params: 128,774,530
Non-trainable params: 0
----------------------------------------------------------------------------
Input size (MB): 9.19
Forward/backward pass size (MB): 2007.94
Params size (MB): 491.24
Estimated Total Size (MB): 2508.36
----------------------------------------------------------------------------
{'total_params': 128774530, 'trainable_params': 128774530}
VGGNet11的训练和验证结果
Predict begin...step 10/10 [==============================] - 6ms/step Predict samples: 10[1 1 0 1 0 0 0 1 0 1]
ResNet
大名鼎鼎的残差网络ResNet，残差块的提出使更深的网络模型成为可能，残差块的结构如图所示， 增加了一个旁路连接，就使深层模型的训练成为可能。 ResNet模型的实现重点在于残差块的实现，这里单独定义一个残差块的类，在forward中
def forward(self,x): copy_x=x x=self.conv(x) if self.in_c!=self.out_c: copy_x=self.w(copy_x) return x+copy_x
分为两条路计算，一条是经典的卷积层，一条是旁路连接，同时如果卷积前后特征图通道数不同，则在旁路中补充一个卷积以调整。
In [16]
class Resblock(nn.Layer):
def __init__(self,in_c,out_c,pool=False):
super(Resblock,self).__init__()
self.in_c,self.out_c=in_c,out_c if pool==True:
self.conv=nn.Sequential(nn.Conv2D(in_c,out_c,3,stride=2,padding=1),nn.Conv2D(out_c,out_c,3,padding=1),nn.BatchNorm(out_c))
self.w=nn.Conv2D(in_c,out_c,kernel_size=1,stride=2) else:
self.conv=nn.Sequential(nn.Conv2D(in_c,out_c,3,padding=1),nn.Conv2D(out_c,out_c,3,padding=1),nn.BatchNorm(out_c))
self.w=nn.Conv2D(in_c,out_c,kernel_size=1)
def forward(self,x):
copy_x=x
x=self.conv(x) if self.in_c!=self.out_c:
copy_x=self.w(copy_x) return x+copy_x
class ResNet(nn.Layer):
def __init__(self):
super().__init__()
self.layers=nn.LayerList()
self.layers.append(nn.Sequential(nn.Conv2D(in_channels=3,out_channels=64,kernel_size=7,stride=2,padding=1)))#,nn.MaxPool2D(2)
for ind in range(3):
self.layers.append(Resblock(64,64))
self.layers.append(Resblock(64,128,True)) for ind in range(3):
self.layers.append(Resblock(128,128))
self.layers.append(Resblock(128,256,True)) for ind in range(5):
self.layers.append(Resblock(256,256))
self.layers.append(Resblock(256,512,True)) for ind in range(2):
self.layers.append(Resblock(512,512))
self.layers.append(nn.Sequential(nn.Flatten(),nn.Linear(100352,2)))
def forward(self,x):
for layer in self.layers:
x=layer(x) return x
network=ResNet()
paddle.summary(network,(16,3,224,224))#使用paddle API验证结构正确性
--------------------------------------------------------------------------- Layer (type) Input Shape Output Shape Param # =========================================================================== Conv2D-59 [[16, 3, 224, 224]] [16, 64, 110, 110] 9,472 Conv2D-60 [[16, 64, 110, 110]] [16, 64, 110, 110] 36,928 Conv2D-61 [[16, 64, 110, 110]] [16, 64, 110, 110] 36,928 BatchNorm-1 [[16, 64, 110, 110]] [16, 64, 110, 110] 256 Resblock-1 [[16, 64, 110, 110]] [16, 64, 110, 110] 0 Conv2D-63 [[16, 64, 110, 110]] [16, 64, 110, 110] 36,928 Conv2D-64 [[16, 64, 110, 110]] [16, 64, 110, 110] 36,928 BatchNorm-2 [[16, 64, 110, 110]] [16, 64, 110, 110] 256 Resblock-2 [[16, 64, 110, 110]] [16, 64, 110, 110] 0 Conv2D-66 [[16, 64, 110, 110]] [16, 64, 110, 110] 36,928 Conv2D-67 [[16, 64, 110, 110]] [16, 64, 110, 110] 36,928 BatchNorm-3 [[16, 64, 110, 110]] [16, 64, 110, 110] 256 Resblock-3 [[16, 64, 110, 110]] [16, 64, 110, 110] 0 Conv2D-69 [[16, 64, 110, 110]] [16, 128, 55, 55] 73,856 Conv2D-70 [[16, 128, 55, 55]] [16, 128, 55, 55] 147,584 BatchNorm-4 [[16, 128, 55, 55]] [16, 128, 55, 55] 512 Conv2D-71 [[16, 64, 110, 110]] [16, 128, 55, 55] 8,320 Resblock-4 [[16, 64, 110, 110]] [16, 128, 55, 55] 0 Conv2D-72 [[16, 128, 55, 55]] [16, 128, 55, 55] 147,584 Conv2D-73 [[16, 128, 55, 55]] [16, 128, 55, 55] 147,584 BatchNorm-5 [[16, 128, 55, 55]] [16, 128, 55, 55] 512 Resblock-5 [[16, 128, 55, 55]] [16, 128, 55, 55] 0 Conv2D-75 [[16, 128, 55, 55]] [16, 128, 55, 55] 147,584 Conv2D-76 [[16, 128, 55, 55]] [16, 128, 55, 55] 147,584 BatchNorm-6 [[16, 128, 55, 55]] [16, 128, 55, 55] 512 Resblock-6 [[16, 128, 55, 55]] [16, 128, 55, 55] 0 Conv2D-78 [[16, 128, 55, 55]] [16, 128, 55, 55] 147,584 Conv2D-79 [[16, 128, 55, 55]] [16, 128, 55, 55] 147,584 BatchNorm-7 [[16, 128, 55, 55]] [16, 128, 55, 55] 512 Resblock-7 [[16, 128, 55, 55]] [16, 128, 55, 55] 0 Conv2D-81 [[16, 128, 55, 55]] [16, 256, 28, 28] 295,168 Conv2D-82 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 BatchNorm-8 [[16, 256, 28, 28]] [16, 256, 28, 28] 1,024 Conv2D-83 [[16, 128, 55, 55]] [16, 256, 28, 28] 33,024 Resblock-8 [[16, 128, 55, 55]] [16, 256, 28, 28] 0 Conv2D-84 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 Conv2D-85 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 BatchNorm-9 [[16, 256, 28, 28]] [16, 256, 28, 28] 1,024 Resblock-9 [[16, 256, 28, 28]] [16, 256, 28, 28] 0 Conv2D-87 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 Conv2D-88 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 BatchNorm-10 [[16, 256, 28, 28]] [16, 256, 28, 28] 1,024 Resblock-10 [[16, 256, 28, 28]] [16, 256, 28, 28] 0 Conv2D-90 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 Conv2D-91 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 BatchNorm-11 [[16, 256, 28, 28]] [16, 256, 28, 28] 1,024 Resblock-11 [[16, 256, 28, 28]] [16, 256, 28, 28] 0 Conv2D-93 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 Conv2D-94 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 BatchNorm-12 [[16, 256, 28, 28]] [16, 256, 28, 28] 1,024 Resblock-12 [[16, 256, 28, 28]] [16, 256, 28, 28] 0 Conv2D-96 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 Conv2D-97 [[16, 256, 28, 28]] [16, 256, 28, 28] 590,080 BatchNorm-13 [[16, 256, 28, 28]] [16, 256, 28, 28] 1,024 Resblock-13 [[16, 256, 28, 28]] [16, 256, 28, 28] 0 Conv2D-99 [[16, 256, 28, 28]] [16, 512, 14, 14] 1,180,160 Conv2D-100 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,359,808 BatchNorm-14 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,048 Conv2D-101 [[16, 256, 28, 28]] [16, 512, 14, 14] 131,584 Resblock-14 [[16, 256, 28, 28]] [16, 512, 14, 14] 0 Conv2D-102 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,359,808 Conv2D-103 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,359,808 BatchNorm-15 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,048 Resblock-15 [[16, 512, 14, 14]] [16, 512, 14, 14] 0 Conv2D-105 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,359,808 Conv2D-106 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,359,808 BatchNorm-16 [[16, 512, 14, 14]] [16, 512, 14, 14] 2,048 Resblock-16 [[16, 512, 14, 14]] [16, 512, 14, 14] 0 Flatten-21 [[16, 512, 14, 14]] [16, 100352] 0 Linear-34 [[16, 100352]] [16, 2] 200,706 =========================================================================== Total params: 21,491,970 Trainable params: 21,476,866 Non-trainable params: 15,104 --------------------------------------------------------------------------- Input size (MB): 9.19 Forward/backward pass size (MB): 2816.42 Params size (MB): 81.99 Estimated Total Size (MB): 2907.59 ---------------------------------------------------------------------------
{'total_params': 21491970, 'trainable_params': 21476866}
在模型训练和验证部分，验证结果如下
Predict begin...step 10/10 [==============================] - 13ms/step Predict samples: 10[1 1 0 1 0 0 0 1 0 1]
DenseNet
DenseNet于2017被提出，作者没有根据业内的潮流对网络进行deeper/wider结构的改造，是对feature特征入手，在网络连通的前提下，对所有特征层进行了连接，在核心组件DenseBlock中，这一层的输入是之前所有层的输出，如图所示
在此基础上，作者构造了DenseNet，其结构如下
In [17]
#DenseNetclass DenseLayer(nn.Layer):
def __init__(self,in_c,growth_rate,bn_size):
super().__init__()
out_c = growth_rate * bn_size
self.layers=nn.Sequential(nn.BatchNorm2D(in_c),nn.ReLU(),nn.Conv2D(in_c,out_c,1),nn.BatchNorm2D(out_c),nn.ReLU(),nn.Conv2D(out_c,growth_rate,3,padding=1)) def forward(self,x):
y = self.layers(x) return yclass DenseBlock(nn.Layer):
def __init__(self,num_layers,in_c,growth_rate,bn_size):
super().__init__()
self.layers=nn.LayerList() for ind in range(num_layers):
self.layers.append(
DenseLayer(in_c=in_c+ind*growth_rate,
growth_rate=growth_rate,
bn_size=bn_size)
)
def forward(self,x):
features=[x] for layer in self.layers:
new_x=layer(paddle.concat(features,axis=1))
features.append(new_x) return paddle.concat(features,axis=1)class Transition(nn.Layer):
def __init__(self,in_c,out_c):
super().__init__()
self.layers=nn.Sequential(nn.BatchNorm2D(in_c),nn.ReLU(),nn.Conv2D(in_c,out_c,1),nn.AvgPool2D(2,2))
def forward(self,x):
return self.layers(x)class DenseNet(nn.Layer):
def __init__(self,num_classes,growth_rate=32,block=(6,12,24,16),bn_size=4,out_c=64):
super().__init__()
self.conv_pool=nn.Sequential(nn.Conv2D(3,out_c,7,stride=2,padding=3),nn.MaxPool2D(3,2))
self.blocks=nn.LayerList()
in_c=out_c for ind,n in enumerate(block):
self.blocks.append(DenseBlock(n,in_c,growth_rate,bn_size))
in_c+=growth_rate*n if ind!=len(block)-1:
self.blocks.append(Transition(in_c,in_c//2))
in_c//=2
self.blocks.append(nn.Sequential(nn.BatchNorm2D(in_c),nn.ReLU(),nn.AdaptiveAvgPool2D((1,1)),nn.Flatten()))
self.cls=nn.Linear(in_c,num_classes)
def forward(self,x):
x=self.conv_pool(x) for layer in self.blocks:
x=layer(x)
x=self.cls(x) return xdef _DenseNet(arch, block_cfg, batch_norm, pretrained, **kwargs):
model = DenseNet(block=block_cfg,**kwargs) if pretrained: assert arch in model_urls, "{} model do not have a pretrained model now, you should set pretrained=False".format(
arch)
weight_path = get_weights_path_from_url(model_urls[arch][0],
model_urls[arch][1])
param = paddle.load(weight_path)
model.load_dict(param) return modeldef DenseNet121(pretrained=False, batch_norm=False, **kwargs):
model_name = 'DenseNet121'
if batch_norm:
model_name += ('_bn') return _DenseNet(model_name, (6,12,24,16), batch_norm, pretrained, **kwargs)def DenseNet161(pretrained=False, batch_norm=False, **kwargs):
model_name = 'DenseNet161'
if batch_norm:
model_name += ('_bn') return _DenseNet(model_name, (6,12,32,32), batch_norm, pretrained, **kwargs)def DenseNet169(pretrained=False, batch_norm=False, **kwargs):
model_name = 'DenseNet169'
if batch_norm:
model_name += ('_bn') return _DenseNet(model_name, (6,12,48,32), batch_norm, pretrained, **kwargs)def DenseNet201(pretrained=False, batch_norm=False, **kwargs):
model_name = 'DenseNet201'
if batch_norm:
model_name += ('_bn') return _DenseNet(model_name, (6,12,64,48), batch_norm, pretrained, **kwargs)
network=DenseNet(2)
paddle.summary(network,(16,3,224,224))#使用paddle API验证结构正确性
-------------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
===============================================================================
Conv2D-108 [[16, 3, 224, 224]] [16, 64, 112, 112] 9,472
MaxPool2D-36 [[16, 64, 112, 112]] [16, 64, 55, 55] 0
BatchNorm2D-1 [[16, 64, 55, 55]] [16, 64, 55, 55] 256
ReLU-81 [[16, 64, 55, 55]] [16, 64, 55, 55] 0
Conv2D-109 [[16, 64, 55, 55]] [16, 128, 55, 55] 8,320
BatchNorm2D-2 [[16, 128, 55, 55]] [16, 128, 55, 55] 512
ReLU-82 [[16, 128, 55, 55]] [16, 128, 55, 55] 0
Conv2D-110 [[16, 128, 55, 55]] [16, 32, 55, 55] 36,896
DenseLayer-1 [[16, 64, 55, 55]] [16, 32, 55, 55] 0
BatchNorm2D-3 [[16, 96, 55, 55]] [16, 96, 55, 55] 384
ReLU-83 [[16, 96, 55, 55]] [16, 96, 55, 55] 0
Conv2D-111 [[16, 96, 55, 55]] [16, 128, 55, 55] 12,416
BatchNorm2D-4 [[16, 128, 55, 55]] [16, 128, 55, 55] 512
ReLU-84 [[16, 128, 55, 55]] [16, 128, 55, 55] 0
Conv2D-112 [[16, 128, 55, 55]] [16, 32, 55, 55] 36,896
DenseLayer-2 [[16, 96, 55, 55]] [16, 32, 55, 55] 0
BatchNorm2D-5 [[16, 128, 55, 55]] [16, 128, 55, 55] 512
ReLU-85 [[16, 128, 55, 55]] [16, 128, 55, 55] 0
Conv2D-113 [[16, 128, 55, 55]] [16, 128, 55, 55] 16,512
BatchNorm2D-6 [[16, 128, 55, 55]] [16, 128, 55, 55] 512
ReLU-86 [[16, 128, 55, 55]] [16, 128, 55, 55] 0
Conv2D-114 [[16, 128, 55, 55]] [16, 32, 55, 55] 36,896
DenseLayer-3 [[16, 128, 55, 55]] [16, 32, 55, 55] 0
BatchNorm2D-7 [[16, 160, 55, 55]] [16, 160, 55, 55] 640
ReLU-87 [[16, 160, 55, 55]] [16, 160, 55, 55] 0
Conv2D-115 [[16, 160, 55, 55]] [16, 128, 55, 55] 20,608
BatchNorm2D-8 [[16, 128, 55, 55]] [16, 128, 55, 55] 512
ReLU-88 [[16, 128, 55, 55]] [16, 128, 55, 55] 0
Conv2D-116 [[16, 128, 55, 55]] [16, 32, 55, 55] 36,896
DenseLayer-4 [[16, 160, 55, 55]] [16, 32, 55, 55] 0
BatchNorm2D-9 [[16, 192, 55, 55]] [16, 192, 55, 55] 768
ReLU-89 [[16, 192, 55, 55]] [16, 192, 55, 55] 0
Conv2D-117 [[16, 192, 55, 55]] [16, 128, 55, 55] 24,704
BatchNorm2D-10 [[16, 128, 55, 55]] [16, 128, 55, 55] 512
ReLU-90 [[16, 128, 55, 55]] [16, 128, 55, 55] 0
Conv2D-118 [[16, 128, 55, 55]] [16, 32, 55, 55] 36,896
DenseLayer-5 [[16, 192, 55, 55]] [16, 32, 55, 55] 0
BatchNorm2D-11 [[16, 224, 55, 55]] [16, 224, 55, 55] 896
ReLU-91 [[16, 224, 55, 55]] [16, 224, 55, 55] 0
Conv2D-119 [[16, 224, 55, 55]] [16, 128, 55, 55] 28,800
BatchNorm2D-12 [[16, 128, 55, 55]] [16, 128, 55, 55] 512
ReLU-92 [[16, 128, 55, 55]] [16, 128, 55, 55] 0
Conv2D-120 [[16, 128, 55, 55]] [16, 32, 55, 55] 36,896
DenseLayer-6 [[16, 224, 55, 55]] [16, 32, 55, 55] 0
DenseBlock-1 [[16, 64, 55, 55]] [16, 256, 55, 55] 0
BatchNorm2D-13 [[16, 256, 55, 55]] [16, 256, 55, 55] 1,024
ReLU-93 [[16, 256, 55, 55]] [16, 256, 55, 55] 0
Conv2D-121 [[16, 256, 55, 55]] [16, 128, 55, 55] 32,896
AvgPool2D-1 [[16, 128, 55, 55]] [16, 128, 27, 27] 0
Transition-1 [[16, 256, 55, 55]] [16, 128, 27, 27] 0
BatchNorm2D-14 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-94 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-122 [[16, 128, 27, 27]] [16, 128, 27, 27] 16,512
BatchNorm2D-15 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-95 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-123 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-7 [[16, 128, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-16 [[16, 160, 27, 27]] [16, 160, 27, 27] 640
ReLU-96 [[16, 160, 27, 27]] [16, 160, 27, 27] 0
Conv2D-124 [[16, 160, 27, 27]] [16, 128, 27, 27] 20,608
BatchNorm2D-17 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-97 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-125 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-8 [[16, 160, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-18 [[16, 192, 27, 27]] [16, 192, 27, 27] 768
ReLU-98 [[16, 192, 27, 27]] [16, 192, 27, 27] 0
Conv2D-126 [[16, 192, 27, 27]] [16, 128, 27, 27] 24,704
BatchNorm2D-19 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-99 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-127 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-9 [[16, 192, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-20 [[16, 224, 27, 27]] [16, 224, 27, 27] 896
ReLU-100 [[16, 224, 27, 27]] [16, 224, 27, 27] 0
Conv2D-128 [[16, 224, 27, 27]] [16, 128, 27, 27] 28,800
BatchNorm2D-21 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-101 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-129 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-10 [[16, 224, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-22 [[16, 256, 27, 27]] [16, 256, 27, 27] 1,024
ReLU-102 [[16, 256, 27, 27]] [16, 256, 27, 27] 0
Conv2D-130 [[16, 256, 27, 27]] [16, 128, 27, 27] 32,896
BatchNorm2D-23 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-103 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-131 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-11 [[16, 256, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-24 [[16, 288, 27, 27]] [16, 288, 27, 27] 1,152
ReLU-104 [[16, 288, 27, 27]] [16, 288, 27, 27] 0
Conv2D-132 [[16, 288, 27, 27]] [16, 128, 27, 27] 36,992
BatchNorm2D-25 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-105 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-133 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-12 [[16, 288, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-26 [[16, 320, 27, 27]] [16, 320, 27, 27] 1,280
ReLU-106 [[16, 320, 27, 27]] [16, 320, 27, 27] 0
Conv2D-134 [[16, 320, 27, 27]] [16, 128, 27, 27] 41,088
BatchNorm2D-27 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-107 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-135 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-13 [[16, 320, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-28 [[16, 352, 27, 27]] [16, 352, 27, 27] 1,408
ReLU-108 [[16, 352, 27, 27]] [16, 352, 27, 27] 0
Conv2D-136 [[16, 352, 27, 27]] [16, 128, 27, 27] 45,184
BatchNorm2D-29 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-109 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-137 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-14 [[16, 352, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-30 [[16, 384, 27, 27]] [16, 384, 27, 27] 1,536
ReLU-110 [[16, 384, 27, 27]] [16, 384, 27, 27] 0
Conv2D-138 [[16, 384, 27, 27]] [16, 128, 27, 27] 49,280
BatchNorm2D-31 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-111 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-139 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-15 [[16, 384, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-32 [[16, 416, 27, 27]] [16, 416, 27, 27] 1,664
ReLU-112 [[16, 416, 27, 27]] [16, 416, 27, 27] 0
Conv2D-140 [[16, 416, 27, 27]] [16, 128, 27, 27] 53,376
BatchNorm2D-33 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-113 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-141 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-16 [[16, 416, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-34 [[16, 448, 27, 27]] [16, 448, 27, 27] 1,792
ReLU-114 [[16, 448, 27, 27]] [16, 448, 27, 27] 0
Conv2D-142 [[16, 448, 27, 27]] [16, 128, 27, 27] 57,472
BatchNorm2D-35 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-115 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-143 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-17 [[16, 448, 27, 27]] [16, 32, 27, 27] 0
BatchNorm2D-36 [[16, 480, 27, 27]] [16, 480, 27, 27] 1,920
ReLU-116 [[16, 480, 27, 27]] [16, 480, 27, 27] 0
Conv2D-144 [[16, 480, 27, 27]] [16, 128, 27, 27] 61,568
BatchNorm2D-37 [[16, 128, 27, 27]] [16, 128, 27, 27] 512
ReLU-117 [[16, 128, 27, 27]] [16, 128, 27, 27] 0
Conv2D-145 [[16, 128, 27, 27]] [16, 32, 27, 27] 36,896
DenseLayer-18 [[16, 480, 27, 27]] [16, 32, 27, 27] 0
DenseBlock-2 [[16, 128, 27, 27]] [16, 512, 27, 27] 0
BatchNorm2D-38 [[16, 512, 27, 27]] [16, 512, 27, 27] 2,048
ReLU-118 [[16, 512, 27, 27]] [16, 512, 27, 27] 0
Conv2D-146 [[16, 512, 27, 27]] [16, 256, 27, 27] 131,328
AvgPool2D-2 [[16, 256, 27, 27]] [16, 256, 13, 13] 0
Transition-2 [[16, 512, 27, 27]] [16, 256, 13, 13] 0
BatchNorm2D-39 [[16, 256, 13, 13]] [16, 256, 13, 13] 1,024
ReLU-119 [[16, 256, 13, 13]] [16, 256, 13, 13] 0
Conv2D-147 [[16, 256, 13, 13]] [16, 128, 13, 13] 32,896
BatchNorm2D-40 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-120 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-148 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-19 [[16, 256, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-41 [[16, 288, 13, 13]] [16, 288, 13, 13] 1,152
ReLU-121 [[16, 288, 13, 13]] [16, 288, 13, 13] 0
Conv2D-149 [[16, 288, 13, 13]] [16, 128, 13, 13] 36,992
BatchNorm2D-42 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-122 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-150 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-20 [[16, 288, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-43 [[16, 320, 13, 13]] [16, 320, 13, 13] 1,280
ReLU-123 [[16, 320, 13, 13]] [16, 320, 13, 13] 0
Conv2D-151 [[16, 320, 13, 13]] [16, 128, 13, 13] 41,088
BatchNorm2D-44 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-124 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-152 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-21 [[16, 320, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-45 [[16, 352, 13, 13]] [16, 352, 13, 13] 1,408
ReLU-125 [[16, 352, 13, 13]] [16, 352, 13, 13] 0
Conv2D-153 [[16, 352, 13, 13]] [16, 128, 13, 13] 45,184
BatchNorm2D-46 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-126 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-154 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-22 [[16, 352, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-47 [[16, 384, 13, 13]] [16, 384, 13, 13] 1,536
ReLU-127 [[16, 384, 13, 13]] [16, 384, 13, 13] 0
Conv2D-155 [[16, 384, 13, 13]] [16, 128, 13, 13] 49,280
BatchNorm2D-48 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-128 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-156 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-23 [[16, 384, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-49 [[16, 416, 13, 13]] [16, 416, 13, 13] 1,664
ReLU-129 [[16, 416, 13, 13]] [16, 416, 13, 13] 0
Conv2D-157 [[16, 416, 13, 13]] [16, 128, 13, 13] 53,376
BatchNorm2D-50 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-130 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-158 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-24 [[16, 416, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-51 [[16, 448, 13, 13]] [16, 448, 13, 13] 1,792
ReLU-131 [[16, 448, 13, 13]] [16, 448, 13, 13] 0
Conv2D-159 [[16, 448, 13, 13]] [16, 128, 13, 13] 57,472
BatchNorm2D-52 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-132 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-160 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-25 [[16, 448, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-53 [[16, 480, 13, 13]] [16, 480, 13, 13] 1,920
ReLU-133 [[16, 480, 13, 13]] [16, 480, 13, 13] 0
Conv2D-161 [[16, 480, 13, 13]] [16, 128, 13, 13] 61,568
BatchNorm2D-54 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-134 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-162 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-26 [[16, 480, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-55 [[16, 512, 13, 13]] [16, 512, 13, 13] 2,048
ReLU-135 [[16, 512, 13, 13]] [16, 512, 13, 13] 0
Conv2D-163 [[16, 512, 13, 13]] [16, 128, 13, 13] 65,664
BatchNorm2D-56 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-136 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-164 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-27 [[16, 512, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-57 [[16, 544, 13, 13]] [16, 544, 13, 13] 2,176
ReLU-137 [[16, 544, 13, 13]] [16, 544, 13, 13] 0
Conv2D-165 [[16, 544, 13, 13]] [16, 128, 13, 13] 69,760
BatchNorm2D-58 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-138 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-166 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-28 [[16, 544, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-59 [[16, 576, 13, 13]] [16, 576, 13, 13] 2,304
ReLU-139 [[16, 576, 13, 13]] [16, 576, 13, 13] 0
Conv2D-167 [[16, 576, 13, 13]] [16, 128, 13, 13] 73,856
BatchNorm2D-60 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-140 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-168 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-29 [[16, 576, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-61 [[16, 608, 13, 13]] [16, 608, 13, 13] 2,432
ReLU-141 [[16, 608, 13, 13]] [16, 608, 13, 13] 0
Conv2D-169 [[16, 608, 13, 13]] [16, 128, 13, 13] 77,952
BatchNorm2D-62 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-142 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-170 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-30 [[16, 608, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-63 [[16, 640, 13, 13]] [16, 640, 13, 13] 2,560
ReLU-143 [[16, 640, 13, 13]] [16, 640, 13, 13] 0
Conv2D-171 [[16, 640, 13, 13]] [16, 128, 13, 13] 82,048
BatchNorm2D-64 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-144 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-172 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-31 [[16, 640, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-65 [[16, 672, 13, 13]] [16, 672, 13, 13] 2,688
ReLU-145 [[16, 672, 13, 13]] [16, 672, 13, 13] 0
Conv2D-173 [[16, 672, 13, 13]] [16, 128, 13, 13] 86,144
BatchNorm2D-66 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-146 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-174 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-32 [[16, 672, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-67 [[16, 704, 13, 13]] [16, 704, 13, 13] 2,816
ReLU-147 [[16, 704, 13, 13]] [16, 704, 13, 13] 0
Conv2D-175 [[16, 704, 13, 13]] [16, 128, 13, 13] 90,240
BatchNorm2D-68 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-148 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-176 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-33 [[16, 704, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-69 [[16, 736, 13, 13]] [16, 736, 13, 13] 2,944
ReLU-149 [[16, 736, 13, 13]] [16, 736, 13, 13] 0
Conv2D-177 [[16, 736, 13, 13]] [16, 128, 13, 13] 94,336
BatchNorm2D-70 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-150 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-178 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-34 [[16, 736, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-71 [[16, 768, 13, 13]] [16, 768, 13, 13] 3,072
ReLU-151 [[16, 768, 13, 13]] [16, 768, 13, 13] 0
Conv2D-179 [[16, 768, 13, 13]] [16, 128, 13, 13] 98,432
BatchNorm2D-72 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-152 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-180 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-35 [[16, 768, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-73 [[16, 800, 13, 13]] [16, 800, 13, 13] 3,200
ReLU-153 [[16, 800, 13, 13]] [16, 800, 13, 13] 0
Conv2D-181 [[16, 800, 13, 13]] [16, 128, 13, 13] 102,528
BatchNorm2D-74 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-154 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-182 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-36 [[16, 800, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-75 [[16, 832, 13, 13]] [16, 832, 13, 13] 3,328
ReLU-155 [[16, 832, 13, 13]] [16, 832, 13, 13] 0
Conv2D-183 [[16, 832, 13, 13]] [16, 128, 13, 13] 106,624
BatchNorm2D-76 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-156 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-184 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-37 [[16, 832, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-77 [[16, 864, 13, 13]] [16, 864, 13, 13] 3,456
ReLU-157 [[16, 864, 13, 13]] [16, 864, 13, 13] 0
Conv2D-185 [[16, 864, 13, 13]] [16, 128, 13, 13] 110,720
BatchNorm2D-78 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-158 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-186 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-38 [[16, 864, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-79 [[16, 896, 13, 13]] [16, 896, 13, 13] 3,584
ReLU-159 [[16, 896, 13, 13]] [16, 896, 13, 13] 0
Conv2D-187 [[16, 896, 13, 13]] [16, 128, 13, 13] 114,816
BatchNorm2D-80 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-160 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-188 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-39 [[16, 896, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-81 [[16, 928, 13, 13]] [16, 928, 13, 13] 3,712
ReLU-161 [[16, 928, 13, 13]] [16, 928, 13, 13] 0
Conv2D-189 [[16, 928, 13, 13]] [16, 128, 13, 13] 118,912
BatchNorm2D-82 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-162 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-190 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-40 [[16, 928, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-83 [[16, 960, 13, 13]] [16, 960, 13, 13] 3,840
ReLU-163 [[16, 960, 13, 13]] [16, 960, 13, 13] 0
Conv2D-191 [[16, 960, 13, 13]] [16, 128, 13, 13] 123,008
BatchNorm2D-84 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-164 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-192 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-41 [[16, 960, 13, 13]] [16, 32, 13, 13] 0
BatchNorm2D-85 [[16, 992, 13, 13]] [16, 992, 13, 13] 3,968
ReLU-165 [[16, 992, 13, 13]] [16, 992, 13, 13] 0
Conv2D-193 [[16, 992, 13, 13]] [16, 128, 13, 13] 127,104
BatchNorm2D-86 [[16, 128, 13, 13]] [16, 128, 13, 13] 512
ReLU-166 [[16, 128, 13, 13]] [16, 128, 13, 13] 0
Conv2D-194 [[16, 128, 13, 13]] [16, 32, 13, 13] 36,896
DenseLayer-42 [[16, 992, 13, 13]] [16, 32, 13, 13] 0
DenseBlock-3 [[16, 256, 13, 13]] [16, 1024, 13, 13] 0
BatchNorm2D-87 [[16, 1024, 13, 13]] [16, 1024, 13, 13] 4,096
ReLU-167 [[16, 1024, 13, 13]] [16, 1024, 13, 13] 0
Conv2D-195 [[16, 1024, 13, 13]] [16, 512, 13, 13] 524,800
AvgPool2D-3 [[16, 512, 13, 13]] [16, 512, 6, 6] 0
Transition-3 [[16, 1024, 13, 13]] [16, 512, 6, 6] 0
BatchNorm2D-88 [[16, 512, 6, 6]] [16, 512, 6, 6] 2,048
ReLU-168 [[16, 512, 6, 6]] [16, 512, 6, 6] 0
Conv2D-196 [[16, 512, 6, 6]] [16, 128, 6, 6] 65,664
BatchNorm2D-89 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-169 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-197 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-43 [[16, 512, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-90 [[16, 544, 6, 6]] [16, 544, 6, 6] 2,176
ReLU-170 [[16, 544, 6, 6]] [16, 544, 6, 6] 0
Conv2D-198 [[16, 544, 6, 6]] [16, 128, 6, 6] 69,760
BatchNorm2D-91 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-171 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-199 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-44 [[16, 544, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-92 [[16, 576, 6, 6]] [16, 576, 6, 6] 2,304
ReLU-172 [[16, 576, 6, 6]] [16, 576, 6, 6] 0
Conv2D-200 [[16, 576, 6, 6]] [16, 128, 6, 6] 73,856
BatchNorm2D-93 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-173 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-201 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-45 [[16, 576, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-94 [[16, 608, 6, 6]] [16, 608, 6, 6] 2,432
ReLU-174 [[16, 608, 6, 6]] [16, 608, 6, 6] 0
Conv2D-202 [[16, 608, 6, 6]] [16, 128, 6, 6] 77,952
BatchNorm2D-95 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-175 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-203 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-46 [[16, 608, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-96 [[16, 640, 6, 6]] [16, 640, 6, 6] 2,560
ReLU-176 [[16, 640, 6, 6]] [16, 640, 6, 6] 0
Conv2D-204 [[16, 640, 6, 6]] [16, 128, 6, 6] 82,048
BatchNorm2D-97 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-177 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-205 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-47 [[16, 640, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-98 [[16, 672, 6, 6]] [16, 672, 6, 6] 2,688
ReLU-178 [[16, 672, 6, 6]] [16, 672, 6, 6] 0
Conv2D-206 [[16, 672, 6, 6]] [16, 128, 6, 6] 86,144
BatchNorm2D-99 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-179 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-207 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-48 [[16, 672, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-100 [[16, 704, 6, 6]] [16, 704, 6, 6] 2,816
ReLU-180 [[16, 704, 6, 6]] [16, 704, 6, 6] 0
Conv2D-208 [[16, 704, 6, 6]] [16, 128, 6, 6] 90,240
BatchNorm2D-101 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-181 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-209 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-49 [[16, 704, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-102 [[16, 736, 6, 6]] [16, 736, 6, 6] 2,944
ReLU-182 [[16, 736, 6, 6]] [16, 736, 6, 6] 0
Conv2D-210 [[16, 736, 6, 6]] [16, 128, 6, 6] 94,336
BatchNorm2D-103 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-183 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-211 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-50 [[16, 736, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-104 [[16, 768, 6, 6]] [16, 768, 6, 6] 3,072
ReLU-184 [[16, 768, 6, 6]] [16, 768, 6, 6] 0
Conv2D-212 [[16, 768, 6, 6]] [16, 128, 6, 6] 98,432
BatchNorm2D-105 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-185 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-213 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-51 [[16, 768, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-106 [[16, 800, 6, 6]] [16, 800, 6, 6] 3,200
ReLU-186 [[16, 800, 6, 6]] [16, 800, 6, 6] 0
Conv2D-214 [[16, 800, 6, 6]] [16, 128, 6, 6] 102,528
BatchNorm2D-107 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-187 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-215 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-52 [[16, 800, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-108 [[16, 832, 6, 6]] [16, 832, 6, 6] 3,328
ReLU-188 [[16, 832, 6, 6]] [16, 832, 6, 6] 0
Conv2D-216 [[16, 832, 6, 6]] [16, 128, 6, 6] 106,624
BatchNorm2D-109 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-189 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-217 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-53 [[16, 832, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-110 [[16, 864, 6, 6]] [16, 864, 6, 6] 3,456
ReLU-190 [[16, 864, 6, 6]] [16, 864, 6, 6] 0
Conv2D-218 [[16, 864, 6, 6]] [16, 128, 6, 6] 110,720
BatchNorm2D-111 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-191 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-219 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-54 [[16, 864, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-112 [[16, 896, 6, 6]] [16, 896, 6, 6] 3,584
ReLU-192 [[16, 896, 6, 6]] [16, 896, 6, 6] 0
Conv2D-220 [[16, 896, 6, 6]] [16, 128, 6, 6] 114,816
BatchNorm2D-113 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-193 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-221 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-55 [[16, 896, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-114 [[16, 928, 6, 6]] [16, 928, 6, 6] 3,712
ReLU-194 [[16, 928, 6, 6]] [16, 928, 6, 6] 0
Conv2D-222 [[16, 928, 6, 6]] [16, 128, 6, 6] 118,912
BatchNorm2D-115 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-195 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-223 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-56 [[16, 928, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-116 [[16, 960, 6, 6]] [16, 960, 6, 6] 3,840
ReLU-196 [[16, 960, 6, 6]] [16, 960, 6, 6] 0
Conv2D-224 [[16, 960, 6, 6]] [16, 128, 6, 6] 123,008
BatchNorm2D-117 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-197 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-225 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-57 [[16, 960, 6, 6]] [16, 32, 6, 6] 0
BatchNorm2D-118 [[16, 992, 6, 6]] [16, 992, 6, 6] 3,968
ReLU-198 [[16, 992, 6, 6]] [16, 992, 6, 6] 0
Conv2D-226 [[16, 992, 6, 6]] [16, 128, 6, 6] 127,104
BatchNorm2D-119 [[16, 128, 6, 6]] [16, 128, 6, 6] 512
ReLU-199 [[16, 128, 6, 6]] [16, 128, 6, 6] 0
Conv2D-227 [[16, 128, 6, 6]] [16, 32, 6, 6] 36,896
DenseLayer-58 [[16, 992, 6, 6]] [16, 32, 6, 6] 0
DenseBlock-4 [[16, 512, 6, 6]] [16, 1024, 6, 6] 0
BatchNorm2D-120 [[16, 1024, 6, 6]] [16, 1024, 6, 6] 4,096
ReLU-200 [[16, 1024, 6, 6]] [16, 1024, 6, 6] 0
AdaptiveAvgPool2D-1 [[16, 1024, 6, 6]] [16, 1024, 1, 1] 0
Flatten-23 [[16, 1024, 1, 1]] [16, 1024] 0
Linear-35 [[16, 1024]] [16, 2] 2,050
===============================================================================
Total params: 7,049,538
Trainable params: 6,882,498
Non-trainable params: 167,040
-------------------------------------------------------------------------------
Input size (MB): 9.19
Forward/backward pass size (MB): 4472.80
Params size (MB): 26.89
Estimated Total Size (MB): 4508.88
-------------------------------------------------------------------------------
{'total_params': 7049538, 'trainable_params': 6882498}
模型训练
In [17]
model=paddle.Model(network) lr=paddle.optimizer.lr.CosineAnnealingDecay(learning_rate=lr_base,T_max=epochs) opt=paddle.optimizer.Momentum(learning_rate=lr,parameters=model.parameters(),weight_decay=1e-2) opt=paddle.optimizer.Adam(learning_rate=lr,parameters=model.parameters()) loss=paddle.nn.CrossEntropyLoss()#axis=1model.prepare(opt, loss,metrics=paddle.metric.Accuracy()) model.fit(train_loader, val_loader, epochs=epochs,verbose=2,save_dir="./net_params",log_freq=1)
The loss value printed in the log is the current step, and the metric is the average value of previous steps. Epoch 1/10
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:641: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance.")
step 1/40 - loss: 0.8786 - acc: 0.0000e+00 - 884ms/step step 2/40 - loss: 1.3529 - acc: 0.3750 - 636ms/step step 3/40 - loss: 4.8848 - acc: 0.3333 - 548ms/step step 4/40 - loss: 0.6855 - acc: 0.4375 - 504ms/step step 5/40 - loss: 1.3280 - acc: 0.4500 - 477ms/step step 6/40 - loss: 1.4665 - acc: 0.4583 - 459ms/step step 7/40 - loss: 0.2446 - acc: 0.5000 - 446ms/step step 8/40 - loss: 0.1795 - acc: 0.5625 - 437ms/step step 9/40 - loss: 0.1281 - acc: 0.6111 - 430ms/step step 10/40 - loss: 0.7968 - acc: 0.6250 - 423ms/step step 11/40 - loss: 2.5774 - acc: 0.5682 - 419ms/step step 12/40 - loss: 0.2166 - acc: 0.6042 - 417ms/step step 13/40 - loss: 2.4652 - acc: 0.5577 - 415ms/step step 14/40 - loss: 1.1314 - acc: 0.5357 - 413ms/step step 15/40 - loss: 0.6019 - acc: 0.5500 - 411ms/step step 16/40 - loss: 1.0493 - acc: 0.5469 - 410ms/step step 17/40 - loss: 0.3270 - acc: 0.5588 - 409ms/step step 18/40 - loss: 0.1344 - acc: 0.5833 - 408ms/step step 19/40 - loss: 0.9306 - acc: 0.5921 - 407ms/step step 20/40 - loss: 0.0057 - acc: 0.6125 - 405ms/step step 21/40 - loss: 0.1379 - acc: 0.6310 - 403ms/step step 22/40 - loss: 0.0118 - acc: 0.6477 - 402ms/step step 23/40 - loss: 0.0922 - acc: 0.6630 - 402ms/step step 24/40 - loss: 0.0295 - acc: 0.6771 - 401ms/step step 25/40 - loss: 0.8993 - acc: 0.6700 - 400ms/step step 26/40 - loss: 0.6457 - acc: 0.6731 - 400ms/step step 27/40 - loss: 0.0824 - acc: 0.6852 - 399ms/step step 28/40 - loss: 0.1024 - acc: 0.6964 - 398ms/step step 29/40 - loss: 0.1232 - acc: 0.7069 - 398ms/step step 30/40 - loss: 0.5631 - acc: 0.7083 - 397ms/step step 31/40 - loss: 0.0370 - acc: 0.7177 - 397ms/step step 32/40 - loss: 0.2733 - acc: 0.7188 - 397ms/step step 33/40 - loss: 6.6403 - acc: 0.7045 - 396ms/step step 34/40 - loss: 0.1885 - acc: 0.7132 - 395ms/step step 35/40 - loss: 0.0506 - acc: 0.7214 - 395ms/step step 36/40 - loss: 3.8153 - acc: 0.7222 - 394ms/step step 37/40 - loss: 5.5249 - acc: 0.7162 - 393ms/step step 38/40 - loss: 0.5504 - acc: 0.7171 - 393ms/step step 39/40 - loss: 0.9939 - acc: 0.7179 - 392ms/step step 40/40 - loss: 0.9726 - acc: 0.7188 - 391ms/step save checkpoint at /home/aistudio/net_params/0 Eval begin... step 1/10 - loss: 0.8339 - acc: 0.7500 - 189ms/step step 2/10 - loss: 3.0489 - acc: 0.5000 - 173ms/step step 3/10 - loss: 6.1806 - acc: 0.4167 - 167ms/step step 4/10 - loss: 2.5625 - acc: 0.4375 - 165ms/step step 5/10 - loss: 6.9693 - acc: 0.4500 - 164ms/step step 6/10 - loss: 7.5170 - acc: 0.4167 - 163ms/step step 7/10 - loss: 2.3508 - acc: 0.3929 - 166ms/step step 8/10 - loss: 4.7935 - acc: 0.4375 - 165ms/step step 9/10 - loss: 9.8368 - acc: 0.3889 - 164ms/step step 10/10 - loss: 4.2736 - acc: 0.4000 - 163ms/step Eval samples: 40 Epoch 2/10 step 1/40 - loss: 0.1939 - acc: 1.0000 - 427ms/step step 2/40 - loss: 0.1233 - acc: 1.0000 - 405ms/step step 3/40 - loss: 0.1640 - acc: 1.0000 - 402ms/step step 4/40 - loss: 0.2072 - acc: 1.0000 - 399ms/step step 5/40 - loss: 0.0291 - acc: 1.0000 - 396ms/step step 6/40 - loss: 0.5622 - acc: 0.9583 - 393ms/step step 7/40 - loss: 1.7429 - acc: 0.8929 - 391ms/step step 8/40 - loss: 0.0600 - acc: 0.9062 - 390ms/step step 9/40 - loss: 0.0239 - acc: 0.9167 - 390ms/step step 10/40 - loss: 0.0170 - acc: 0.9250 - 388ms/step step 11/40 - loss: 0.4259 - acc: 0.9091 - 387ms/step step 12/40 - loss: 0.3709 - acc: 0.8958 - 386ms/step step 13/40 - loss: 0.2406 - acc: 0.9038 - 384ms/step step 14/40 - loss: 0.0292 - acc: 0.9107 - 384ms/step step 15/40 - loss: 0.8347 - acc: 0.9000 - 384ms/step step 16/40 - loss: 0.0132 - acc: 0.9062 - 384ms/step step 17/40 - loss: 0.1719 - acc: 0.9118 - 384ms/step step 18/40 - loss: 0.1203 - acc: 0.9167 - 384ms/step step 19/40 - loss: 0.1806 - acc: 0.9211 - 383ms/step step 20/40 - loss: 0.0590 - acc: 0.9250 - 383ms/step step 21/40 - loss: 0.7904 - acc: 0.9167 - 383ms/step step 22/40 - loss: 0.7270 - acc: 0.9091 - 383ms/step step 23/40 - loss: 0.0995 - acc: 0.9130 - 382ms/step step 24/40 - loss: 0.6964 - acc: 0.9062 - 383ms/step step 25/40 - loss: 0.0703 - acc: 0.9100 - 382ms/step step 26/40 - loss: 0.2966 - acc: 0.9038 - 382ms/step step 27/40 - loss: 0.0068 - acc: 0.9074 - 382ms/step step 28/40 - loss: 3.3363 - acc: 0.9018 - 382ms/step step 29/40 - loss: 0.0243 - acc: 0.9052 - 383ms/step step 30/40 - loss: 0.0606 - acc: 0.9083 - 383ms/step step 31/40 - loss: 0.5678 - acc: 0.9032 - 383ms/step step 32/40 - loss: 0.0569 - acc: 0.9062 - 383ms/step step 33/40 - loss: 0.0124 - acc: 0.9091 - 383ms/step step 34/40 - loss: 0.0041 - acc: 0.9118 - 383ms/step step 35/40 - loss: 0.1410 - acc: 0.9143 - 383ms/step step 36/40 - loss: 0.0177 - acc: 0.9167 - 382ms/step step 37/40 - loss: 0.4238 - acc: 0.9122 - 382ms/step step 38/40 - loss: 0.0226 - acc: 0.9145 - 382ms/step step 39/40 - loss: 0.7144 - acc: 0.9103 - 381ms/step step 40/40 - loss: 0.0323 - acc: 0.9125 - 381ms/step save checkpoint at /home/aistudio/net_params/1 Eval begin... step 1/10 - loss: 1.5320 - acc: 0.7500 - 190ms/step step 2/10 - loss: 0.0654 - acc: 0.8750 - 173ms/step step 3/10 - loss: 0.1166 - acc: 0.9167 - 167ms/step step 4/10 - loss: 0.0883 - acc: 0.9375 - 164ms/step step 5/10 - loss: 0.0637 - acc: 0.9500 - 162ms/step step 6/10 - loss: 0.7424 - acc: 0.9167 - 160ms/step step 7/10 - loss: 0.1827 - acc: 0.9286 - 159ms/step step 8/10 - loss: 0.0102 - acc: 0.9375 - 159ms/step step 9/10 - loss: 8.4945e-04 - acc: 0.9444 - 159ms/step step 10/10 - loss: 0.0039 - acc: 0.9500 - 158ms/step Eval samples: 40 Epoch 3/10 step 1/40 - loss: 0.0395 - acc: 1.0000 - 401ms/step step 2/40 - loss: 0.1322 - acc: 1.0000 - 388ms/step step 3/40 - loss: 0.0062 - acc: 1.0000 - 386ms/step step 4/40 - loss: 0.0462 - acc: 1.0000 - 387ms/step step 5/40 - loss: 1.5052 - acc: 0.9000 - 386ms/step step 6/40 - loss: 0.0362 - acc: 0.9167 - 384ms/step step 7/40 - loss: 0.1628 - acc: 0.9286 - 383ms/step step 8/40 - loss: 0.1661 - acc: 0.9375 - 383ms/step step 9/40 - loss: 1.4469 - acc: 0.9167 - 382ms/step step 10/40 - loss: 0.4381 - acc: 0.9000 - 382ms/step step 11/40 - loss: 0.4622 - acc: 0.8864 - 381ms/step step 12/40 - loss: 1.9192 - acc: 0.8542 - 380ms/step step 13/40 - loss: 0.1092 - acc: 0.8654 - 380ms/step step 14/40 - loss: 0.0249 - acc: 0.8750 - 380ms/step step 15/40 - loss: 0.1569 - acc: 0.8833 - 381ms/step step 16/40 - loss: 0.6889 - acc: 0.8750 - 381ms/step step 17/40 - loss: 1.1773 - acc: 0.8676 - 384ms/step step 18/40 - loss: 0.1330 - acc: 0.8750 - 384ms/step step 19/40 - loss: 0.0544 - acc: 0.8816 - 383ms/step step 20/40 - loss: 0.1582 - acc: 0.8875 - 383ms/step step 21/40 - loss: 0.1629 - acc: 0.8929 - 383ms/step step 22/40 - loss: 0.2393 - acc: 0.8864 - 382ms/step step 23/40 - loss: 0.2353 - acc: 0.8913 - 382ms/step step 24/40 - loss: 0.1751 - acc: 0.8958 - 382ms/step step 25/40 - loss: 0.0879 - acc: 0.9000 - 382ms/step step 26/40 - loss: 0.0663 - acc: 0.9038 - 382ms/step step 27/40 - loss: 1.2867 - acc: 0.8981 - 382ms/step step 28/40 - loss: 0.0812 - acc: 0.9018 - 382ms/step step 29/40 - loss: 0.0434 - acc: 0.9052 - 382ms/step step 30/40 - loss: 0.9762 - acc: 0.9000 - 382ms/step step 31/40 - loss: 0.9838 - acc: 0.8952 - 382ms/step step 32/40 - loss: 0.0421 - acc: 0.8984 - 382ms/step step 33/40 - loss: 0.0318 - acc: 0.9015 - 382ms/step step 34/40 - loss: 0.0415 - acc: 0.9044 - 382ms/step step 35/40 - loss: 0.7238 - acc: 0.9000 - 382ms/step step 36/40 - loss: 0.0645 - acc: 0.9028 - 382ms/step step 37/40 - loss: 0.2908 - acc: 0.8986 - 382ms/step step 38/40 - loss: 0.0585 - acc: 0.9013 - 381ms/step step 39/40 - loss: 0.1431 - acc: 0.9038 - 381ms/step step 40/40 - loss: 0.2628 - acc: 0.9062 - 381ms/step save checkpoint at /home/aistudio/net_params/2 Eval begin... step 1/10 - loss: 0.6657 - acc: 0.7500 - 189ms/step step 2/10 - loss: 0.2851 - acc: 0.8750 - 173ms/step step 3/10 - loss: 0.1932 - acc: 0.9167 - 168ms/step step 4/10 - loss: 0.1568 - acc: 0.9375 - 165ms/step step 5/10 - loss: 0.2661 - acc: 0.9500 - 163ms/step step 6/10 - loss: 0.3660 - acc: 0.9167 - 162ms/step step 7/10 - loss: 0.1076 - acc: 0.9286 - 161ms/step step 8/10 - loss: 0.3483 - acc: 0.9062 - 160ms/step step 9/10 - loss: 0.0626 - acc: 0.9167 - 159ms/step step 10/10 - loss: 2.5330 - acc: 0.9000 - 158ms/step Eval samples: 40 Epoch 4/10 step 1/40 - loss: 0.0737 - acc: 1.0000 - 410ms/step step 2/40 - loss: 0.1885 - acc: 1.0000 - 393ms/step step 3/40 - loss: 0.2421 - acc: 1.0000 - 388ms/step step 4/40 - loss: 0.3262 - acc: 1.0000 - 388ms/step step 5/40 - loss: 0.1046 - acc: 1.0000 - 387ms/step step 6/40 - loss: 0.3956 - acc: 0.9583 - 386ms/step step 7/40 - loss: 0.0820 - acc: 0.9643 - 384ms/step step 8/40 - loss: 0.0214 - acc: 0.9688 - 383ms/step step 9/40 - loss: 2.7114 - acc: 0.9167 - 384ms/step step 10/40 - loss: 0.3407 - acc: 0.9000 - 383ms/step step 11/40 - loss: 0.1124 - acc: 0.9091 - 383ms/step step 12/40 - loss: 0.0579 - acc: 0.9167 - 383ms/step step 13/40 - loss: 2.1480 - acc: 0.8846 - 383ms/step step 14/40 - loss: 1.6564 - acc: 0.8393 - 385ms/step step 15/40 - loss: 2.2699 - acc: 0.8167 - 390ms/step step 16/40 - loss: 0.0320 - acc: 0.8281 - 390ms/step step 17/40 - loss: 0.0675 - acc: 0.8382 - 390ms/step step 18/40 - loss: 0.7838 - acc: 0.8333 - 391ms/step step 19/40 - loss: 0.8567 - acc: 0.8289 - 391ms/step step 20/40 - loss: 1.2302 - acc: 0.8125 - 392ms/step step 21/40 - loss: 0.8249 - acc: 0.8095 - 392ms/step step 22/40 - loss: 0.0762 - acc: 0.8182 - 391ms/step step 23/40 - loss: 0.6335 - acc: 0.8152 - 391ms/step step 24/40 - loss: 1.4432 - acc: 0.8125 - 390ms/step step 25/40 - loss: 0.3524 - acc: 0.8100 - 390ms/step step 26/40 - loss: 0.0775 - acc: 0.8173 - 389ms/step step 27/40 - loss: 0.0042 - acc: 0.8241 - 389ms/step step 28/40 - loss: 0.4740 - acc: 0.8214 - 389ms/step step 29/40 - loss: 0.0156 - acc: 0.8276 - 389ms/step step 30/40 - loss: 0.7265 - acc: 0.8167 - 389ms/step step 31/40 - loss: 0.0151 - acc: 0.8226 - 389ms/step step 32/40 - loss: 2.1609 - acc: 0.8047 - 389ms/step step 33/40 - loss: 0.7140 - acc: 0.7955 - 388ms/step step 34/40 - loss: 0.9281 - acc: 0.7794 - 388ms/step step 35/40 - loss: 0.3517 - acc: 0.7786 - 388ms/step step 36/40 - loss: 0.2083 - acc: 0.7778 - 387ms/step step 37/40 - loss: 0.0585 - acc: 0.7838 - 387ms/step step 38/40 - loss: 0.0416 - acc: 0.7895 - 386ms/step step 39/40 - loss: 0.0436 - acc: 0.7949 - 386ms/step step 40/40 - loss: 0.2235 - acc: 0.7937 - 386ms/step save checkpoint at /home/aistudio/net_params/3 Eval begin... step 1/10 - loss: 0.4787 - acc: 0.5000 - 193ms/step step 2/10 - loss: 0.2900 - acc: 0.7500 - 176ms/step step 3/10 - loss: 0.3595 - acc: 0.7500 - 169ms/step step 4/10 - loss: 1.2558 - acc: 0.7500 - 166ms/step step 5/10 - loss: 0.1971 - acc: 0.8000 - 164ms/step step 6/10 - loss: 0.2971 - acc: 0.7917 - 162ms/step step 7/10 - loss: 0.3777 - acc: 0.7857 - 160ms/step step 8/10 - loss: 0.3725 - acc: 0.8125 - 159ms/step step 9/10 - loss: 0.0464 - acc: 0.8333 - 159ms/step step 10/10 - loss: 0.1397 - acc: 0.8500 - 159ms/step Eval samples: 40 Epoch 5/10 step 1/40 - loss: 0.0323 - acc: 1.0000 - 419ms/step step 2/40 - loss: 0.6679 - acc: 0.8750 - 409ms/step step 3/40 - loss: 1.4082 - acc: 0.7500 - 405ms/step step 4/40 - loss: 0.1318 - acc: 0.8125 - 403ms/step step 5/40 - loss: 0.0193 - acc: 0.8500 - 401ms/step step 6/40 - loss: 0.2944 - acc: 0.8333 - 397ms/step step 7/40 - loss: 6.1585 - acc: 0.7857 - 394ms/step step 8/40 - loss: 0.5659 - acc: 0.7500 - 392ms/step step 9/40 - loss: 0.7713 - acc: 0.7500 - 390ms/step step 10/40 - loss: 1.2466 - acc: 0.7250 - 389ms/step step 11/40 - loss: 5.7933 - acc: 0.7045 - 388ms/step step 12/40 - loss: 0.0074 - acc: 0.7292 - 388ms/step step 13/40 - loss: 0.3855 - acc: 0.7308 - 389ms/step step 14/40 - loss: 0.2039 - acc: 0.7500 - 389ms/step step 15/40 - loss: 0.1062 - acc: 0.7667 - 388ms/step step 16/40 - loss: 0.0897 - acc: 0.7812 - 388ms/step step 17/40 - loss: 0.0531 - acc: 0.7941 - 388ms/step step 18/40 - loss: 0.0057 - acc: 0.8056 - 387ms/step step 19/40 - loss: 3.8333 - acc: 0.7763 - 387ms/step step 20/40 - loss: 9.8705e-04 - acc: 0.7875 - 387ms/step step 21/40 - loss: 0.0078 - acc: 0.7976 - 387ms/step step 22/40 - loss: 0.1706 - acc: 0.8068 - 387ms/step step 23/40 - loss: 0.0057 - acc: 0.8152 - 387ms/step step 24/40 - loss: 0.7214 - acc: 0.8125 - 387ms/step step 25/40 - loss: 2.2940 - acc: 0.8000 - 387ms/step step 26/40 - loss: 0.1119 - acc: 0.8077 - 387ms/step step 27/40 - loss: 0.1626 - acc: 0.8148 - 387ms/step step 28/40 - loss: 0.1642 - acc: 0.8214 - 387ms/step step 29/40 - loss: 1.4720 - acc: 0.8190 - 386ms/step step 30/40 - loss: 0.1344 - acc: 0.8250 - 386ms/step step 31/40 - loss: 0.0357 - acc: 0.8306 - 386ms/step step 32/40 - loss: 0.9776 - acc: 0.8281 - 386ms/step step 33/40 - loss: 0.0590 - acc: 0.8333 - 385ms/step step 34/40 - loss: 0.0067 - acc: 0.8382 - 385ms/step step 35/40 - loss: 0.0716 - acc: 0.8429 - 385ms/step step 36/40 - loss: 1.0589 - acc: 0.8403 - 385ms/step step 37/40 - loss: 1.0528 - acc: 0.8243 - 385ms/step step 38/40 - loss: 0.1098 - acc: 0.8289 - 385ms/step step 39/40 - loss: 1.0890 - acc: 0.8141 - 385ms/step step 40/40 - loss: 0.0311 - acc: 0.8187 - 385ms/step save checkpoint at /home/aistudio/net_params/4 Eval begin... step 1/10 - loss: 0.6312 - acc: 0.7500 - 190ms/step step 2/10 - loss: 0.0878 - acc: 0.8750 - 174ms/step step 3/10 - loss: 0.3535 - acc: 0.8333 - 167ms/step step 4/10 - loss: 0.6739 - acc: 0.8125 - 164ms/step step 5/10 - loss: 0.4551 - acc: 0.8000 - 162ms/step step 6/10 - loss: 0.2485 - acc: 0.8333 - 161ms/step step 7/10 - loss: 0.0581 - acc: 0.8571 - 159ms/step step 8/10 - loss: 1.1049 - acc: 0.8438 - 158ms/step step 9/10 - loss: 0.7219 - acc: 0.8333 - 157ms/step step 10/10 - loss: 0.0683 - acc: 0.8500 - 157ms/step Eval samples: 40 Epoch 6/10 step 1/40 - loss: 0.0995 - acc: 1.0000 - 409ms/step step 2/40 - loss: 0.0347 - acc: 1.0000 - 393ms/step step 3/40 - loss: 0.0175 - acc: 1.0000 - 387ms/step step 4/40 - loss: 0.0142 - acc: 1.0000 - 384ms/step step 5/40 - loss: 0.0089 - acc: 1.0000 - 382ms/step step 6/40 - loss: 0.0130 - acc: 1.0000 - 382ms/step step 7/40 - loss: 0.0065 - acc: 1.0000 - 381ms/step step 8/40 - loss: 0.0886 - acc: 1.0000 - 381ms/step step 9/40 - loss: 0.0057 - acc: 1.0000 - 382ms/step step 10/40 - loss: 0.4009 - acc: 0.9750 - 383ms/step step 11/40 - loss: 0.0063 - acc: 0.9773 - 383ms/step step 12/40 - loss: 0.1222 - acc: 0.9792 - 385ms/step step 13/40 - loss: 0.2443 - acc: 0.9615 - 384ms/step step 14/40 - loss: 0.0042 - acc: 0.9643 - 383ms/step step 15/40 - loss: 0.1865 - acc: 0.9500 - 383ms/step step 16/40 - loss: 0.2686 - acc: 0.9375 - 384ms/step step 17/40 - loss: 0.0906 - acc: 0.9412 - 384ms/step step 18/40 - loss: 2.0408 - acc: 0.9028 - 383ms/step step 19/40 - loss: 2.7789 - acc: 0.8816 - 383ms/step step 20/40 - loss: 1.5114 - acc: 0.8625 - 382ms/step step 21/40 - loss: 0.0674 - acc: 0.8690 - 382ms/step step 22/40 - loss: 0.2582 - acc: 0.8750 - 382ms/step step 23/40 - loss: 0.2616 - acc: 0.8804 - 382ms/step step 24/40 - loss: 0.9854 - acc: 0.8750 - 382ms/step step 25/40 - loss: 0.5641 - acc: 0.8700 - 383ms/step step 26/40 - loss: 0.2137 - acc: 0.8654 - 383ms/step step 27/40 - loss: 1.1425 - acc: 0.8519 - 383ms/step step 28/40 - loss: 2.1342 - acc: 0.8214 - 383ms/step step 29/40 - loss: 0.1785 - acc: 0.8276 - 383ms/step step 30/40 - loss: 0.1765 - acc: 0.8333 - 383ms/step step 31/40 - loss: 0.3265 - acc: 0.8306 - 383ms/step step 32/40 - loss: 0.1236 - acc: 0.8359 - 383ms/step step 33/40 - loss: 0.0458 - acc: 0.8409 - 382ms/step step 34/40 - loss: 0.1007 - acc: 0.8456 - 383ms/step step 35/40 - loss: 0.0287 - acc: 0.8500 - 382ms/step step 36/40 - loss: 0.6687 - acc: 0.8403 - 383ms/step step 37/40 - loss: 0.1288 - acc: 0.8446 - 383ms/step step 38/40 - loss: 0.3275 - acc: 0.8421 - 383ms/step step 39/40 - loss: 0.0299 - acc: 0.8462 - 383ms/step step 40/40 - loss: 0.0082 - acc: 0.8500 - 382ms/step save checkpoint at /home/aistudio/net_params/5 Eval begin... step 1/10 - loss: 0.6713 - acc: 0.7500 - 190ms/step step 2/10 - loss: 0.0364 - acc: 0.8750 - 173ms/step step 3/10 - loss: 0.1046 - acc: 0.9167 - 167ms/step step 4/10 - loss: 0.0863 - acc: 0.9375 - 164ms/step step 5/10 - loss: 0.0156 - acc: 0.9500 - 162ms/step step 6/10 - loss: 0.3814 - acc: 0.9167 - 160ms/step step 7/10 - loss: 0.1301 - acc: 0.9286 - 159ms/step step 8/10 - loss: 0.1887 - acc: 0.9375 - 158ms/step step 9/10 - loss: 5.1071e-04 - acc: 0.9444 - 157ms/step step 10/10 - loss: 0.0056 - acc: 0.9500 - 156ms/step Eval samples: 40 Epoch 7/10 step 1/40 - loss: 0.0386 - acc: 1.0000 - 405ms/step step 2/40 - loss: 0.0222 - acc: 1.0000 - 390ms/step step 3/40 - loss: 0.0136 - acc: 1.0000 - 385ms/step step 4/40 - loss: 0.0808 - acc: 1.0000 - 383ms/step step 5/40 - loss: 0.2800 - acc: 0.9500 - 382ms/step step 6/40 - loss: 3.6075 - acc: 0.8750 - 382ms/step step 7/40 - loss: 0.0278 - acc: 0.8929 - 381ms/step step 8/40 - loss: 0.2057 - acc: 0.9062 - 382ms/step step 9/40 - loss: 0.0269 - acc: 0.9167 - 383ms/step step 10/40 - loss: 0.0028 - acc: 0.9250 - 384ms/step step 11/40 - loss: 0.7834 - acc: 0.9091 - 384ms/step step 12/40 - loss: 0.5458 - acc: 0.8958 - 384ms/step step 13/40 - loss: 0.1274 - acc: 0.9038 - 383ms/step step 14/40 - loss: 0.0356 - acc: 0.9107 - 383ms/step step 15/40 - loss: 0.0309 - acc: 0.9167 - 383ms/step step 16/40 - loss: 0.0779 - acc: 0.9219 - 384ms/step step 17/40 - loss: 0.2377 - acc: 0.9118 - 384ms/step step 18/40 - loss: 0.0049 - acc: 0.9167 - 385ms/step step 19/40 - loss: 0.2741 - acc: 0.9079 - 386ms/step step 20/40 - loss: 0.0322 - acc: 0.9125 - 386ms/step step 21/40 - loss: 0.7266 - acc: 0.9048 - 387ms/step step 22/40 - loss: 0.0301 - acc: 0.9091 - 387ms/step step 23/40 - loss: 0.0418 - acc: 0.9130 - 390ms/step step 24/40 - loss: 0.0183 - acc: 0.9167 - 390ms/step step 25/40 - loss: 0.0116 - acc: 0.9200 - 389ms/step step 26/40 - loss: 0.0014 - acc: 0.9231 - 390ms/step step 27/40 - loss: 0.0010 - acc: 0.9259 - 389ms/step step 28/40 - loss: 0.0080 - acc: 0.9286 - 389ms/step step 29/40 - loss: 0.8775 - acc: 0.9224 - 389ms/step step 30/40 - loss: 1.5853 - acc: 0.9167 - 389ms/step step 31/40 - loss: 0.0105 - acc: 0.9194 - 389ms/step step 32/40 - loss: 0.1602 - acc: 0.9219 - 389ms/step step 33/40 - loss: 0.0230 - acc: 0.9242 - 388ms/step step 34/40 - loss: 0.0125 - acc: 0.9265 - 388ms/step step 35/40 - loss: 0.9842 - acc: 0.9214 - 388ms/step step 36/40 - loss: 0.3961 - acc: 0.9167 - 388ms/step step 37/40 - loss: 0.0035 - acc: 0.9189 - 388ms/step step 38/40 - loss: 0.0021 - acc: 0.9211 - 388ms/step step 39/40 - loss: 0.0378 - acc: 0.9231 - 388ms/step step 40/40 - loss: 0.0681 - acc: 0.9250 - 388ms/step save checkpoint at /home/aistudio/net_params/6 Eval begin... step 1/10 - loss: 1.1088 - acc: 0.7500 - 191ms/step step 2/10 - loss: 0.0657 - acc: 0.8750 - 175ms/step step 3/10 - loss: 0.3504 - acc: 0.8333 - 169ms/step step 4/10 - loss: 0.3822 - acc: 0.8125 - 166ms/step step 5/10 - loss: 0.0095 - acc: 0.8500 - 164ms/step step 6/10 - loss: 0.1725 - acc: 0.8750 - 163ms/step step 7/10 - loss: 0.0494 - acc: 0.8929 - 162ms/step step 8/10 - loss: 0.0705 - acc: 0.9062 - 162ms/step step 9/10 - loss: 0.0017 - acc: 0.9167 - 162ms/step step 10/10 - loss: 0.1170 - acc: 0.9250 - 161ms/step Eval samples: 40 Epoch 8/10 step 1/40 - loss: 0.0471 - acc: 1.0000 - 414ms/step step 2/40 - loss: 0.1036 - acc: 1.0000 - 405ms/step step 3/40 - loss: 0.0255 - acc: 1.0000 - 398ms/step step 4/40 - loss: 0.0952 - acc: 1.0000 - 396ms/step step 5/40 - loss: 0.0220 - acc: 1.0000 - 392ms/step step 6/40 - loss: 0.0714 - acc: 1.0000 - 390ms/step step 7/40 - loss: 0.1415 - acc: 1.0000 - 388ms/step step 8/40 - loss: 0.0573 - acc: 1.0000 - 387ms/step step 9/40 - loss: 0.4687 - acc: 0.9722 - 388ms/step step 10/40 - loss: 0.0601 - acc: 0.9750 - 388ms/step step 11/40 - loss: 3.3628 - acc: 0.9318 - 388ms/step step 12/40 - loss: 0.1056 - acc: 0.9375 - 387ms/step step 13/40 - loss: 0.0047 - acc: 0.9423 - 387ms/step step 14/40 - loss: 0.0695 - acc: 0.9464 - 386ms/step step 15/40 - loss: 0.0321 - acc: 0.9500 - 385ms/step step 16/40 - loss: 0.0046 - acc: 0.9531 - 385ms/step step 17/40 - loss: 0.2005 - acc: 0.9559 - 385ms/step step 18/40 - loss: 0.0053 - acc: 0.9583 - 384ms/step step 19/40 - loss: 2.9423 - acc: 0.9342 - 384ms/step step 20/40 - loss: 0.0326 - acc: 0.9375 - 384ms/step step 21/40 - loss: 0.0439 - acc: 0.9405 - 384ms/step step 22/40 - loss: 0.3001 - acc: 0.9318 - 386ms/step step 23/40 - loss: 0.4708 - acc: 0.9239 - 386ms/step step 24/40 - loss: 0.1299 - acc: 0.9271 - 386ms/step step 25/40 - loss: 0.3625 - acc: 0.9200 - 385ms/step step 26/40 - loss: 0.3287 - acc: 0.9135 - 385ms/step step 27/40 - loss: 0.0549 - acc: 0.9167 - 385ms/step step 28/40 - loss: 0.3235 - acc: 0.9107 - 384ms/step step 29/40 - loss: 0.0694 - acc: 0.9138 - 384ms/step step 30/40 - loss: 0.0509 - acc: 0.9167 - 384ms/step step 31/40 - loss: 0.0484 - acc: 0.9194 - 384ms/step step 32/40 - loss: 0.0848 - acc: 0.9219 - 384ms/step step 33/40 - loss: 0.8888 - acc: 0.9091 - 384ms/step step 34/40 - loss: 0.1689 - acc: 0.9118 - 383ms/step step 35/40 - loss: 0.1079 - acc: 0.9143 - 384ms/step step 36/40 - loss: 1.2262 - acc: 0.8958 - 384ms/step step 37/40 - loss: 0.0499 - acc: 0.8986 - 384ms/step step 38/40 - loss: 0.0687 - acc: 0.9013 - 383ms/step step 39/40 - loss: 0.1622 - acc: 0.9038 - 383ms/step step 40/40 - loss: 1.3731 - acc: 0.8938 - 383ms/step save checkpoint at /home/aistudio/net_params/7 Eval begin... step 1/10 - loss: 0.6370 - acc: 0.7500 - 189ms/step step 2/10 - loss: 0.1336 - acc: 0.8750 - 173ms/step step 3/10 - loss: 0.2266 - acc: 0.8333 - 166ms/step step 4/10 - loss: 0.6489 - acc: 0.8125 - 163ms/step step 5/10 - loss: 0.0253 - acc: 0.8500 - 161ms/step step 6/10 - loss: 0.1293 - acc: 0.8750 - 160ms/step step 7/10 - loss: 0.1190 - acc: 0.8929 - 158ms/step step 8/10 - loss: 0.1791 - acc: 0.9062 - 157ms/step step 9/10 - loss: 0.0681 - acc: 0.9167 - 157ms/step step 10/10 - loss: 0.2486 - acc: 0.9000 - 156ms/step Eval samples: 40 Epoch 9/10 step 1/40 - loss: 0.1473 - acc: 1.0000 - 412ms/step step 2/40 - loss: 0.0333 - acc: 1.0000 - 395ms/step step 3/40 - loss: 0.1590 - acc: 1.0000 - 392ms/step step 4/40 - loss: 0.1155 - acc: 1.0000 - 388ms/step step 5/40 - loss: 0.1285 - acc: 1.0000 - 385ms/step step 6/40 - loss: 0.0225 - acc: 1.0000 - 383ms/step step 7/40 - loss: 0.0925 - acc: 1.0000 - 382ms/step step 8/40 - loss: 0.0555 - acc: 1.0000 - 383ms/step step 9/40 - loss: 0.1942 - acc: 1.0000 - 383ms/step step 10/40 - loss: 0.1221 - acc: 1.0000 - 385ms/step step 11/40 - loss: 0.1720 - acc: 1.0000 - 385ms/step step 12/40 - loss: 0.1029 - acc: 1.0000 - 385ms/step step 13/40 - loss: 0.3487 - acc: 0.9808 - 385ms/step step 14/40 - loss: 0.0366 - acc: 0.9821 - 384ms/step step 15/40 - loss: 0.1332 - acc: 0.9833 - 384ms/step step 16/40 - loss: 0.1053 - acc: 0.9844 - 383ms/step step 17/40 - loss: 0.8791 - acc: 0.9706 - 384ms/step step 18/40 - loss: 0.0092 - acc: 0.9722 - 383ms/step step 19/40 - loss: 0.1165 - acc: 0.9737 - 384ms/step step 20/40 - loss: 0.0231 - acc: 0.9750 - 384ms/step step 21/40 - loss: 0.0517 - acc: 0.9762 - 386ms/step step 22/40 - loss: 1.2441 - acc: 0.9545 - 386ms/step step 23/40 - loss: 0.0475 - acc: 0.9565 - 386ms/step step 24/40 - loss: 0.0250 - acc: 0.9583 - 385ms/step step 25/40 - loss: 1.8250 - acc: 0.9300 - 385ms/step step 26/40 - loss: 1.5704 - acc: 0.9135 - 385ms/step step 27/40 - loss: 0.0336 - acc: 0.9167 - 385ms/step step 28/40 - loss: 0.1095 - acc: 0.9196 - 385ms/step step 29/40 - loss: 0.0183 - acc: 0.9224 - 385ms/step step 30/40 - loss: 1.0537 - acc: 0.9000 - 384ms/step step 31/40 - loss: 0.0366 - acc: 0.9032 - 384ms/step step 32/40 - loss: 0.3328 - acc: 0.8984 - 384ms/step step 33/40 - loss: 1.4123 - acc: 0.8864 - 384ms/step step 34/40 - loss: 0.0306 - acc: 0.8897 - 384ms/step step 35/40 - loss: 0.1041 - acc: 0.8929 - 384ms/step step 36/40 - loss: 0.2017 - acc: 0.8889 - 384ms/step step 37/40 - loss: 0.0764 - acc: 0.8919 - 384ms/step step 38/40 - loss: 0.4936 - acc: 0.8816 - 384ms/step step 39/40 - loss: 0.0707 - acc: 0.8846 - 384ms/step step 40/40 - loss: 0.2644 - acc: 0.8812 - 384ms/step save checkpoint at /home/aistudio/net_params/8 Eval begin... step 1/10 - loss: 0.5388 - acc: 0.7500 - 191ms/step step 2/10 - loss: 0.1467 - acc: 0.8750 - 174ms/step step 3/10 - loss: 0.3554 - acc: 0.8333 - 167ms/step step 4/10 - loss: 0.9144 - acc: 0.8125 - 164ms/step step 5/10 - loss: 0.0390 - acc: 0.8500 - 162ms/step step 6/10 - loss: 0.1069 - acc: 0.8750 - 161ms/step step 7/10 - loss: 0.0991 - acc: 0.8929 - 160ms/step step 8/10 - loss: 0.3669 - acc: 0.8750 - 159ms/step step 9/10 - loss: 0.0068 - acc: 0.8889 - 158ms/step step 10/10 - loss: 0.1100 - acc: 0.9000 - 158ms/step Eval samples: 40 Epoch 10/10 step 1/40 - loss: 0.3692 - acc: 0.7500 - 411ms/step step 2/40 - loss: 0.0414 - acc: 0.8750 - 396ms/step step 3/40 - loss: 0.3528 - acc: 0.8333 - 390ms/step step 4/40 - loss: 1.5622 - acc: 0.6875 - 389ms/step step 5/40 - loss: 1.3839 - acc: 0.6500 - 387ms/step step 6/40 - loss: 0.0628 - acc: 0.7083 - 386ms/step step 7/40 - loss: 0.0734 - acc: 0.7500 - 386ms/step step 8/40 - loss: 0.1471 - acc: 0.7812 - 386ms/step step 9/40 - loss: 0.1939 - acc: 0.7778 - 385ms/step step 10/40 - loss: 0.0424 - acc: 0.8000 - 385ms/step step 11/40 - loss: 0.0248 - acc: 0.8182 - 385ms/step step 12/40 - loss: 0.1522 - acc: 0.8333 - 384ms/step step 13/40 - loss: 0.0541 - acc: 0.8462 - 383ms/step step 14/40 - loss: 0.2955 - acc: 0.8571 - 384ms/step step 15/40 - loss: 0.0316 - acc: 0.8667 - 384ms/step step 16/40 - loss: 0.1238 - acc: 0.8750 - 384ms/step step 17/40 - loss: 0.0661 - acc: 0.8824 - 383ms/step step 18/40 - loss: 1.5378 - acc: 0.8472 - 383ms/step step 19/40 - loss: 0.0806 - acc: 0.8553 - 383ms/step step 20/40 - loss: 1.9089 - acc: 0.8250 - 384ms/step step 21/40 - loss: 0.1667 - acc: 0.8333 - 385ms/step step 22/40 - loss: 0.0567 - acc: 0.8409 - 386ms/step step 23/40 - loss: 0.0559 - acc: 0.8478 - 386ms/step step 24/40 - loss: 0.1250 - acc: 0.8542 - 386ms/step step 25/40 - loss: 0.7509 - acc: 0.8300 - 387ms/step step 26/40 - loss: 0.0650 - acc: 0.8365 - 387ms/step step 27/40 - loss: 0.2121 - acc: 0.8426 - 388ms/step step 28/40 - loss: 0.9168 - acc: 0.8304 - 388ms/step step 29/40 - loss: 0.3841 - acc: 0.8276 - 388ms/step step 30/40 - loss: 0.2089 - acc: 0.8333 - 388ms/step step 31/40 - loss: 0.5309 - acc: 0.8306 - 387ms/step step 32/40 - loss: 0.0864 - acc: 0.8359 - 387ms/step step 33/40 - loss: 0.8717 - acc: 0.8258 - 386ms/step step 34/40 - loss: 0.0905 - acc: 0.8309 - 387ms/step step 35/40 - loss: 0.4231 - acc: 0.8214 - 386ms/step step 36/40 - loss: 0.7544 - acc: 0.8125 - 386ms/step step 37/40 - loss: 0.0584 - acc: 0.8176 - 386ms/step step 38/40 - loss: 0.1380 - acc: 0.8224 - 386ms/step step 39/40 - loss: 0.3090 - acc: 0.8269 - 386ms/step step 40/40 - loss: 0.1033 - acc: 0.8313 - 385ms/step save checkpoint at /home/aistudio/net_params/9 Eval begin... step 1/10 - loss: 0.6522 - acc: 0.7500 - 191ms/step step 2/10 - loss: 0.2859 - acc: 0.8750 - 174ms/step step 3/10 - loss: 0.2603 - acc: 0.8333 - 168ms/step step 4/10 - loss: 0.3827 - acc: 0.8125 - 165ms/step step 5/10 - loss: 0.1566 - acc: 0.8500 - 163ms/step step 6/10 - loss: 0.3308 - acc: 0.8333 - 161ms/step step 7/10 - loss: 0.2705 - acc: 0.8571 - 160ms/step step 8/10 - loss: 0.1988 - acc: 0.8750 - 159ms/step step 9/10 - loss: 0.1806 - acc: 0.8889 - 159ms/step step 10/10 - loss: 0.2267 - acc: 0.8750 - 159ms/step Eval samples: 40 save checkpoint at /home/aistudio/net_params/final
模型验证
In [19]
test_set=DatasetN("test_list.txt",T.Compose([T.Normalize(data_format="CHW")]))
model.load("net_params/final")
test_result=model.predict(test_set)[0]print(np.argmax(test_result,axis=2).flatten())#1 1 0 1 0 0 0 1 0 1
Predict begin... step 10/10 [==============================] - 153ms/step Predict samples: 10 [1 1 1 1 0 0 0 1 1 1]
In [20]
!rm -r net_params/*
以上就是自己动手搭建模型：图像分类篇的详细内容，更多请关注启程网【www.vszh.cn】AI 人工智能栏目。
本文地址：https://www.vszh.cn/article/1159387.html
如非特殊说明，本站内容均来自于网友自主分享，如有任何问题均请联系我们进行处理！
猜您喜欢
失控进化新手电梯搭建指南
2026-07-03失控进化电梯如何搭建，随着家园层数不断增加，爬楼梯上下楼逐渐变成“体力负担”——尤其是搬运大量物资时，来来回回跑不仅耗时，还容易错过宝贵的探索时间。那么电梯的搭建就是必不可少了。 失控进化新手电梯搭建指南： 一、核心材料 自建电梯的核心是“电梯主体+供电系统+控制装置”三大模块： 按实际层数准备电梯模块即可(比如3层就准备3个)组成垂直通行的通道,每层1个,最多可叠加6层。 风力发电机/太阳能板提...
新《VR女友》重磅更新 支持VRM个性模型导入功能
2026-06-18ILLUMINATION社（并非那个I社）宣布，旗下延续I社名作精神“续作”新《VR（的）女友》最新功能上线，重点是加入了对于VRM个性模型的支持，玩家可以与自己喜欢的类型的VR女友进行互动了。 新推出的“VRM女主角功能”只需将VRM格式的3D模型放入指定文件夹并启动，该模型就会作为女主角出现在《VR女友》正作中。 以往的角色只有“夕阳樱花”，但此次功能支持使用VRoid等软件制作和分发的VRM...
-
Anthropic 秘密提交上市申请，最强安全大模型 Mythos 突然放开内测
2026-06-04全球人工智能安全格局正经历技术突破与资本动向的双重冲击。知名AI研究机构Anthropic于本周二正式对外披露，将全面扩展其前沿安全大模型“神话”（Mythos）的全球内测规模，新增约150家组织获得早期访问权限，目标是在正式发布前协同全球关键基础设施运营方开展系统性风险识别与防护加固。 在当前愈演愈烈的AI技术竞备浪潮中，Mythos凭借卓越的代码生成能力与自动化漏洞探测水平，被广泛视为兼具颠覆...
世界模型，挤满了00后
2026-06-04☞☞☞AI智能聊天,问答助手,AI智能搜索,多模态理解力帮你轻松跨越从0到1的创作门槛☜☜☜ 倘若科技创投也存在代际更替，那么这场AI领域的创业浪潮，或许正是史上节奏最急、落差最大、颠覆性最强的一次年轻化跃迁。 没有渐进式演进，没有温和交接，只有一群00后以“破壁者”之姿，强势切入AI最底层、最艰深的战场——“世界模型”。 与移动互联网时代按部就班的成长路径截然不同，这批千禧年后出生的数字原住民，...
大模型如何真正告别幻觉？元认知或是破局的关键
2026-06-04大模型所表现出的“幻觉”——即以高度确信的语气输出明显违背事实的内容——始终是ai领域难以绕开的关键挑战，尤其在医疗、司法等容错率极低的垂直场景中，此类错误可能带来严重后果。 过去，行业主要依赖两种策略应对幻觉：其一，持续扩充训练语料，寄望于让模型覆盖更广的知识边界；其二，部署保守型响应机制，要求模型在置信度不足时主动拒答。但二者均存在根本性瓶颈：前者受限于数据获取与标注的现实约束，注定无法填满所...
-
超越GPT-5.5！国产AI大模型MiniMax M3 正式发布
2026-06-02国内人工智能领域迎来里程碑式技术突破。稀宇科技今日正式推出全新一代大模型——minimaxm3。该模型不仅拥有业界领先的代码生成与理解能力，更原生支持长达1m（即100万tokens）的超长上下文窗口。尤为值得关注的是，m3首次在国内开源模型中实现三大核心能力一体化：原生支持图像、视频多模态输入，并可直接理解与操作电脑桌面界面，标志着国产大模型在多模态智能体方向迈出关键一步。 ☞☞☞AI智能聊天,...
LobsterAI上线图片视频大模型矩阵 一次性接入四大主流图像视频生成模型
2026-06-02☞☞☞AI智能聊天,问答助手,AI智能搜索,多模态理解力帮你轻松跨越从0到1的创作门槛☜☜☜ 国内AIGC多模态创作领域迎来新一轮突破。近日，国内头部互联网企业推出的首个开源“龙虾系”人工智能产品LobsterAI（网易有道龙虾）完成重大迭代，正式开放图片与视频生成两大核心能力。 本次升级引入全新矩阵化协同架构，同步集成业界公认的四大先进多模态生成模型——Seedream、Seedance、Hap...
-
Anthropic发布Claude Opus 4.8模型，强化协作与安全性
2026-06-025月29日讯，anthropic正式推出claudeopus4.8模型，在智能体编程、多学科推理及知识型代理任务等关键能力上相较opus4.7取得实质性跃升，同时延续原有定价策略，实现“加量不加价”。 本次更新重点强化模型的“可信协作”属性，尤其在“诚实度”维度实现突破性优化：其对自身生成代码中缺陷的主动识别与标注能力显著增强，漏过错误而不提示的概率较前代下降约四倍；在对齐性评估中，“隐性欺骗”“...