本项目用paddle2.0复现PyramidNet，在10分类动物数据集上训练验证。该网络通道数随深度逐模块增加，用加法方式及Padding适配通道。对比ResNet，其残差模块层排列不同。实验显示，PyramidNet在动物分类上准确率高于ResNet50，参数更少。还复现了其在Cifar10和Cifar100上的训练验证。
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜
项目背景
PyramidNet是2017年CVPR上的一篇论文Deep Pyramidal Residual Networks中提出的网络模型。由于很多论文中都出现该网络，故本项目即对其进行复现。
卷积网络在计算机视觉领域可以说是一个‘must have’的网络模块，通常，卷积网络是多层叠加的，卷积网络的通道数量只是在feature map的大小发生减小的时候出现增加的变化。而PyramidNet是在每个卷积模块后都会有通道数量的增加，从而形成通道数量逐模块逐渐增加的金字塔结构。论文作者通过实验发现，这种网络结构设计有助于网络的鲁棒性。本项目就来对金字塔网络一探究竟，用PyramidNet来玩一下牛年识牛的动物图像分类实验。
项目简介
本项目首次使用paddle2.0复现了金字塔网络PyramidNet，并在动物数据集上进行了训练和验证。
动物数据集的划分是按8:2的的划分方法进行训练集与验证集划分的。
模型简介
PyramidNet网络的核心思想是随着网络深度的加深，通道的数量逐渐增加。PyramidNet网络和ResNet网络的比较如图1所示。
图1 PyramidNet网络和ResNet网络的结构比较图
既然PyramidNet网络的结构是随着网络深度的加深，通道数量逐层增加，那么增加的方式就有加法和乘法两种方式。图2给出了这两种方式及它们比较的图示。其中a是加法方式，b是乘法方式，c是两种方式的比较。可以看出，b这种方式更接近传统的ResNet网络。论文中通过实验得出加法方式比乘法方式效果要好，故本项目采用加法的通道数量增加方法。
图2 PyramidNet以加法和乘法方式增加通道数量的比较图示
模型细节
由于PyramidNet网络的通道数量是逐模块增加的，它的残差模块在做残差相加的时候就要对通道数量做适配。PyramidNet采取Padding的方式对通道数量做适配，实现细节如图3所示。其中a是通过padding做通道数量适配的示意图，b是通过padding做通道数量适配的等效分解示意图。
图3 PyramidNet残差模块通道数量适配示意图
PyramidNet网络的另一个实现细节是对残差模块的卷积层、BatchNorm层和ReLU的排列组合进行了有别于ResNet的设计，具体实现可以fork后见代码细节。
论文原文：Deep Pyramidal Residual Networks
参考代码：
数据集介绍
本项目使用10分类的动物数据集进行训练和验证.
该十分类动物数据集，包含dog,horse,elephant,butterfly,chicken,cat,cow,sheep,spider和squirrel。每一分类的图片数量为2k-5k。
文件结构
| 文件名或文件夹名 | 功能 |
|---|---|
| PyramidNet.py | 金字塔网络模型文件 |
| animal_dataset.py | 数据集定义文件 |
| config.py | 配置文件 |
| train_val_split.py | 训练集验证集划分文件 |
| train_cifar.py | 复现论文训练cifar数据集的文件 |
| train.py | 模型训练文件 |
| eval.py | 模型验证文件 |
解压数据集
In [ ]
!unzip -q data/data70196/animals.zip -d work/dataset
查看图片
In [ ]
import osimport randomfrom matplotlib import pyplot as pltfrom PIL import Image
imgs = []
paths = os.listdir('work/dataset')for path in paths:
img_path = os.path.join('work/dataset', path) if os.path.isdir(img_path):
img_paths = os.listdir(img_path)
img = Image.open(os.path.join(img_path, random.choice(img_paths)))
imgs.append((img, path))
f, ax = plt.subplots(3, 3, figsize=(12,12))for i, img in enumerate(imgs[:9]):
ax[i//3, i%3].imshow(img[0])
ax[i//3, i%3].axis('off')
ax[i//3, i%3].set_title('label: %s' % img[1])
plt.show()划分训练集和验证集
In [ ]
!python code/train_val_split.py
finished train val split!
打印网络结构
In [7]
%cd code
/home/aistudio/code
In [10]
import paddle
In [11]
from PyramidNet import PyramidNet
model_basic = PyramidNet('imagenet', 32, 300, num_classes=10, bottleneck=False)
model_basic = paddle.Model(model_basic)
model_basic.summary((-1, 3, 224, 224))=> the layer configuration for each stage is set to [3, 3, 3, 3]
---------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
===========================================================================
Conv2D-26 [[1, 3, 224, 224]] [1, 64, 112, 112] 9,408
BatchNorm2D-39 [[1, 64, 112, 112]] [1, 64, 112, 112] 256
ReLU-15 [[1, 64, 112, 112]] [1, 64, 112, 112] 0
MaxPool2D-2 [[1, 64, 112, 112]] [1, 64, 56, 56] 0
BatchNorm2D-40 [[1, 64, 56, 56]] [1, 64, 56, 56] 256
Conv2D-27 [[1, 64, 56, 56]] [1, 89, 56, 56] 51,264
BatchNorm2D-41 [[1, 89, 56, 56]] [1, 89, 56, 56] 356
ReLU-16 [[1, 89, 56, 56]] [1, 89, 56, 56] 0
Conv2D-28 [[1, 89, 56, 56]] [1, 89, 56, 56] 71,289
BatchNorm2D-42 [[1, 89, 56, 56]] [1, 89, 56, 56] 356
BasicBlock-13 [[1, 64, 56, 56]] [1, 89, 56, 56] 0
BatchNorm2D-43 [[1, 89, 56, 56]] [1, 89, 56, 56] 356
Conv2D-29 [[1, 89, 56, 56]] [1, 114, 56, 56] 91,314
BatchNorm2D-44 [[1, 114, 56, 56]] [1, 114, 56, 56] 456
ReLU-17 [[1, 114, 56, 56]] [1, 114, 56, 56] 0
Conv2D-30 [[1, 114, 56, 56]] [1, 114, 56, 56] 116,964
BatchNorm2D-45 [[1, 114, 56, 56]] [1, 114, 56, 56] 456
BasicBlock-14 [[1, 89, 56, 56]] [1, 114, 56, 56] 0
BatchNorm2D-46 [[1, 114, 56, 56]] [1, 114, 56, 56] 456
Conv2D-31 [[1, 114, 56, 56]] [1, 139, 56, 56] 142,614
BatchNorm2D-47 [[1, 139, 56, 56]] [1, 139, 56, 56] 556
ReLU-18 [[1, 139, 56, 56]] [1, 139, 56, 56] 0
Conv2D-32 [[1, 139, 56, 56]] [1, 139, 56, 56] 173,889
BatchNorm2D-48 [[1, 139, 56, 56]] [1, 139, 56, 56] 556
BasicBlock-15 [[1, 114, 56, 56]] [1, 139, 56, 56] 0
BatchNorm2D-49 [[1, 139, 56, 56]] [1, 139, 56, 56] 556
Conv2D-33 [[1, 139, 56, 56]] [1, 164, 28, 28] 205,164
BatchNorm2D-50 [[1, 164, 28, 28]] [1, 164, 28, 28] 656
ReLU-19 [[1, 164, 28, 28]] [1, 164, 28, 28] 0
Conv2D-34 [[1, 164, 28, 28]] [1, 164, 28, 28] 242,064
BatchNorm2D-51 [[1, 164, 28, 28]] [1, 164, 28, 28] 656
AvgPool2D-5 [[1, 139, 56, 56]] [1, 139, 28, 28] 0
BasicBlock-16 [[1, 139, 56, 56]] [1, 164, 28, 28] 0
BatchNorm2D-52 [[1, 164, 28, 28]] [1, 164, 28, 28] 656
Conv2D-35 [[1, 164, 28, 28]] [1, 189, 28, 28] 278,964
BatchNorm2D-53 [[1, 189, 28, 28]] [1, 189, 28, 28] 756
ReLU-20 [[1, 189, 28, 28]] [1, 189, 28, 28] 0
Conv2D-36 [[1, 189, 28, 28]] [1, 189, 28, 28] 321,489
BatchNorm2D-54 [[1, 189, 28, 28]] [1, 189, 28, 28] 756
BasicBlock-17 [[1, 164, 28, 28]] [1, 189, 28, 28] 0
BatchNorm2D-55 [[1, 189, 28, 28]] [1, 189, 28, 28] 756
Conv2D-37 [[1, 189, 28, 28]] [1, 214, 28, 28] 364,014
BatchNorm2D-56 [[1, 214, 28, 28]] [1, 214, 28, 28] 856
ReLU-21 [[1, 214, 28, 28]] [1, 214, 28, 28] 0
Conv2D-38 [[1, 214, 28, 28]] [1, 214, 28, 28] 412,164
BatchNorm2D-57 [[1, 214, 28, 28]] [1, 214, 28, 28] 856
BasicBlock-18 [[1, 189, 28, 28]] [1, 214, 28, 28] 0
BatchNorm2D-58 [[1, 214, 28, 28]] [1, 214, 28, 28] 856
Conv2D-39 [[1, 214, 28, 28]] [1, 239, 14, 14] 460,314
BatchNorm2D-59 [[1, 239, 14, 14]] [1, 239, 14, 14] 956
ReLU-22 [[1, 239, 14, 14]] [1, 239, 14, 14] 0
Conv2D-40 [[1, 239, 14, 14]] [1, 239, 14, 14] 514,089
BatchNorm2D-60 [[1, 239, 14, 14]] [1, 239, 14, 14] 956
AvgPool2D-6 [[1, 214, 28, 28]] [1, 214, 14, 14] 0
BasicBlock-19 [[1, 214, 28, 28]] [1, 239, 14, 14] 0
BatchNorm2D-61 [[1, 239, 14, 14]] [1, 239, 14, 14] 956
Conv2D-41 [[1, 239, 14, 14]] [1, 264, 14, 14] 567,864
BatchNorm2D-62 [[1, 264, 14, 14]] [1, 264, 14, 14] 1,056
ReLU-23 [[1, 264, 14, 14]] [1, 264, 14, 14] 0
Conv2D-42 [[1, 264, 14, 14]] [1, 264, 14, 14] 627,264
BatchNorm2D-63 [[1, 264, 14, 14]] [1, 264, 14, 14] 1,056
BasicBlock-20 [[1, 239, 14, 14]] [1, 264, 14, 14] 0
BatchNorm2D-64 [[1, 264, 14, 14]] [1, 264, 14, 14] 1,056
Conv2D-43 [[1, 264, 14, 14]] [1, 289, 14, 14] 686,664
BatchNorm2D-65 [[1, 289, 14, 14]] [1, 289, 14, 14] 1,156
ReLU-24 [[1, 289, 14, 14]] [1, 289, 14, 14] 0
Conv2D-44 [[1, 289, 14, 14]] [1, 289, 14, 14] 751,689
BatchNorm2D-66 [[1, 289, 14, 14]] [1, 289, 14, 14] 1,156
BasicBlock-21 [[1, 264, 14, 14]] [1, 289, 14, 14] 0
BatchNorm2D-67 [[1, 289, 14, 14]] [1, 289, 14, 14] 1,156
Conv2D-45 [[1, 289, 14, 14]] [1, 314, 7, 7] 816,714
BatchNorm2D-68 [[1, 314, 7, 7]] [1, 314, 7, 7] 1,256
ReLU-25 [[1, 314, 7, 7]] [1, 314, 7, 7] 0
Conv2D-46 [[1, 314, 7, 7]] [1, 314, 7, 7] 887,364
BatchNorm2D-69 [[1, 314, 7, 7]] [1, 314, 7, 7] 1,256
AvgPool2D-7 [[1, 289, 14, 14]] [1, 289, 7, 7] 0
BasicBlock-22 [[1, 289, 14, 14]] [1, 314, 7, 7] 0
BatchNorm2D-70 [[1, 314, 7, 7]] [1, 314, 7, 7] 1,256
Conv2D-47 [[1, 314, 7, 7]] [1, 339, 7, 7] 958,014
BatchNorm2D-71 [[1, 339, 7, 7]] [1, 339, 7, 7] 1,356
ReLU-26 [[1, 339, 7, 7]] [1, 339, 7, 7] 0
Conv2D-48 [[1, 339, 7, 7]] [1, 339, 7, 7] 1,034,289
BatchNorm2D-72 [[1, 339, 7, 7]] [1, 339, 7, 7] 1,356
BasicBlock-23 [[1, 314, 7, 7]] [1, 339, 7, 7] 0
BatchNorm2D-73 [[1, 339, 7, 7]] [1, 339, 7, 7] 1,356
Conv2D-49 [[1, 339, 7, 7]] [1, 364, 7, 7] 1,110,564
BatchNorm2D-74 [[1, 364, 7, 7]] [1, 364, 7, 7] 1,456
ReLU-27 [[1, 364, 7, 7]] [1, 364, 7, 7] 0
Conv2D-50 [[1, 364, 7, 7]] [1, 364, 7, 7] 1,192,464
BatchNorm2D-75 [[1, 364, 7, 7]] [1, 364, 7, 7] 1,456
BasicBlock-24 [[1, 339, 7, 7]] [1, 364, 7, 7] 0
BatchNorm2D-76 [[1, 364, 7, 7]] [1, 364, 7, 7] 1,456
ReLU-28 [[1, 364, 7, 7]] [1, 364, 7, 7] 0
AvgPool2D-8 [[1, 364, 7, 7]] [1, 364, 1, 1] 0
Linear-2 [[1, 364]] [1, 10] 3,650
===========================================================================
Total params: 12,124124,672
Trainable params: 12,091,544
Non-trainable params: 33,128
---------------------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 114.81
Params size (MB): 46.25
Estimated Total Size (MB): 161.63
---------------------------------------------------------------------------{'total_params': 12124124672, 'trainable_params': 12091544}In [12]
from PyramidNet import PyramidNet
model_bottleneck = PyramidNet('imagenet', 32, 300, num_classes=10, bottleneck=True)
model_bottleneck = paddle.Model(model_bottleneck)
model_bottleneck.summary((-1, 3, 224, 224))=> the layer configuration for each stage is set to [2, 2, 2, 2]
---------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
===========================================================================
Conv2D-51 [[1, 3, 224, 224]] [1, 64, 112, 112] 9,408
BatchNorm2D-77 [[1, 64, 112, 112]] [1, 64, 112, 112] 256
ReLU-29 [[1, 64, 112, 112]] [1, 64, 112, 112] 0
MaxPool2D-3 [[1, 64, 112, 112]] [1, 64, 56, 56] 0
BatchNorm2D-78 [[1, 64, 56, 56]] [1, 64, 56, 56] 256
Conv2D-52 [[1, 64, 56, 56]] [1, 102, 56, 56] 6,528
BatchNorm2D-79 [[1, 102, 56, 56]] [1, 102, 56, 56] 408
ReLU-30 [[1, 102, 56, 56]] [1, 102, 56, 56] 0
Conv2D-53 [[1, 102, 56, 56]] [1, 102, 56, 56] 93,636
BatchNorm2D-80 [[1, 102, 56, 56]] [1, 102, 56, 56] 408
Conv2D-54 [[1, 102, 56, 56]] [1, 408, 56, 56] 41,616
BatchNorm2D-81 [[1, 408, 56, 56]] [1, 408, 56, 56] 1,632
Bottleneck-1 [[1, 64, 56, 56]] [1, 408, 56, 56] 0
BatchNorm2D-82 [[1, 408, 56, 56]] [1, 408, 56, 56] 1,632
Conv2D-55 [[1, 408, 56, 56]] [1, 139, 56, 56] 56,712
BatchNorm2D-83 [[1, 139, 56, 56]] [1, 139, 56, 56] 556
ReLU-31 [[1, 139, 56, 56]] [1, 139, 56, 56] 0
Conv2D-56 [[1, 139, 56, 56]] [1, 139, 56, 56] 173,889
BatchNorm2D-84 [[1, 139, 56, 56]] [1, 139, 56, 56] 556
Conv2D-57 [[1, 139, 56, 56]] [1, 556, 56, 56] 77,284
BatchNorm2D-85 [[1, 556, 56, 56]] [1, 556, 56, 56] 2,224
Bottleneck-2 [[1, 408, 56, 56]] [1, 556, 56, 56] 0
BatchNorm2D-86 [[1, 556, 56, 56]] [1, 556, 56, 56] 2,224
Conv2D-58 [[1, 556, 56, 56]] [1, 176, 56, 56] 97,856
BatchNorm2D-87 [[1, 176, 56, 56]] [1, 176, 56, 56] 704
ReLU-32 [[1, 176, 28, 28]] [1, 176, 28, 28] 0
Conv2D-59 [[1, 176, 56, 56]] [1, 176, 28, 28] 278,784
BatchNorm2D-88 [[1, 176, 28, 28]] [1, 176, 28, 28] 704
Conv2D-60 [[1, 176, 28, 28]] [1, 704, 28, 28] 123,904
BatchNorm2D-89 [[1, 704, 28, 28]] [1, 704, 28, 28] 2,816
AvgPool2D-9 [[1, 556, 56, 56]] [1, 556, 28, 28] 0
Bottleneck-3 [[1, 556, 56, 56]] [1, 704, 28, 28] 0
BatchNorm2D-90 [[1, 704, 28, 28]] [1, 704, 28, 28] 2,816
Conv2D-61 [[1, 704, 28, 28]] [1, 214, 28, 28] 150,656
BatchNorm2D-91 [[1, 214, 28, 28]] [1, 214, 28, 28] 856
ReLU-33 [[1, 214, 28, 28]] [1, 214, 28, 28] 0
Conv2D-62 [[1, 214, 28, 28]] [1, 214, 28, 28] 412,164
BatchNorm2D-92 [[1, 214, 28, 28]] [1, 214, 28, 28] 856
Conv2D-63 [[1, 214, 28, 28]] [1, 856, 28, 28] 183,184
BatchNorm2D-93 [[1, 856, 28, 28]] [1, 856, 28, 28] 3,424
Bottleneck-4 [[1, 704, 28, 28]] [1, 856, 28, 28] 0
BatchNorm2D-94 [[1, 856, 28, 28]] [1, 856, 28, 28] 3,424
Conv2D-64 [[1, 856, 28, 28]] [1, 252, 28, 28] 215,712
BatchNorm2D-95 [[1, 252, 28, 28]] [1, 252, 28, 28] 1,008
ReLU-34 [[1, 252, 14, 14]] [1, 252, 14, 14] 0
Conv2D-65 [[1, 252, 28, 28]] [1, 252, 14, 14] 571,536
BatchNorm2D-96 [[1, 252, 14, 14]] [1, 252, 14, 14] 1,008
Conv2D-66 [[1, 252, 14, 14]] [1, 1008, 14, 14] 254,016
BatchNorm2D-97 [[1, 1008, 14, 14]] [1, 1008, 14, 14] 4,032
AvgPool2D-10 [[1, 856, 28, 28]] [1, 856, 14, 14] 0
Bottleneck-5 [[1, 856, 28, 28]] [1, 1008, 14, 14] 0
BatchNorm2D-98 [[1, 1008, 14, 14]] [1, 1008, 14, 14] 4,032
Conv2D-67 [[1, 1008, 14, 14]] [1, 289, 14, 14] 291,312
BatchNorm2D-99 [[1, 289, 14, 14]] [1, 289, 14, 14] 1,156
ReLU-35 [[1, 289, 14, 14]] [1, 289, 14, 14] 0
Conv2D-68 [[1, 289, 14, 14]] [1, 289, 14, 14] 751,689
BatchNorm2D-100 [[1, 289, 14, 14]] [1, 289, 14, 14] 1,156
Conv2D-69 [[1, 289, 14, 14]] [1, 1156, 14, 14] 334,084
BatchNorm2D-101 [[1, 1156, 14, 14]] [1, 1156, 14, 14] 4,624
Bottleneck-6 [[1, 1008, 14, 14]] [1, 1156, 14, 14] 0
BatchNorm2D-102 [[1, 1156, 14, 14]] [1, 1156, 14, 14] 4,624
Conv2D-70 [[1, 1156, 14, 14]] [1, 326, 14, 14] 376,856
BatchNorm2D-103 [[1, 326, 14, 14]] [1, 326, 14, 14] 1,304
ReLU-36 [[1, 326, 7, 7]] [1, 326, 7, 7] 0
Conv2D-71 [[1, 326, 14, 14]] [1, 326, 7, 7] 956,484
BatchNorm2D-104 [[1, 326, 7, 7]] [1, 326, 7, 7] 1,304
Conv2D-72 [[1, 326, 7, 7]] [1, 1304, 7, 7] 425,104
BatchNorm2D-105 [[1, 1304, 7, 7]] [1, 1304, 7, 7] 5,216
AvgPool2D-11 [[1, 1156, 14, 14]] [1, 1156, 7, 7] 0
Bottleneck-7 [[1, 1156, 14, 14]] [1, 1304, 7, 7] 0
BatchNorm2D-106 [[1, 1304, 7, 7]] [1, 1304, 7, 7] 5,216
Conv2D-73 [[1, 1304, 7, 7]] [1, 364, 7, 7] 474,656
BatchNorm2D-107 [[1, 364, 7, 7]] [1, 364, 7, 7] 1,456
ReLU-37 [[1, 364, 7, 7]] [1, 364, 7, 7] 0
Conv2D-74 [[1, 364, 7, 7]] [1, 364, 7, 7] 1,192,464
BatchNorm2D-108 [[1, 364, 7, 7]] [1, 364, 7, 7] 1,456
Conv2D-75 [[1, 364, 7, 7]] [1, 1456, 7, 7] 529,984
BatchNorm2D-109 [[1, 1456, 7, 7]] [1, 1456, 7, 7] 5,824
Bottleneck-8 [[1, 1304, 7, 7]] [1, 1456, 7, 7] 0
BatchNorm2D-110 [[1, 1456, 7, 7]] [1, 1456, 7, 7] 5,824
ReLU-38 [[1, 1456, 7, 7]] [1, 1456, 7, 7] 0
AvgPool2D-12 [[1, 1456, 7, 7]] [1, 1456, 1, 1] 0
Linear-3 [[1, 1456]] [1, 10] 14,570
===========================================================================
Total params: 8,169,080
Trainable params: 8,094,088
Non-trainable params: 74,992
---------------------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 228.82
Params size (MB): 31.16
Estimated Total Size (MB): 260.56
---------------------------------------------------------------------------{'total_params': 8169080, 'trainable_params': 8094088}复现论文原文对Cifar10和Cifar100的训练和验证
首先复现论文中对Cifar10和Cifar100的训练和验证，超参数采用论文中的值。
In [ ]
!python train_cifar.py --net 'pyramidnet_bottleneck' --dataset 'cifar10' --alpha 200 --depth 272 --num_classes 10 --epochs 300 --lr 0.1 --batch_size 128
Eval samples: 10000
图示cifar10训练验证过程
图4. 使用PyramidNet_bottleneck对cifar10的训练验证图示
In [ ]
!python train_cifar.py --net 'pyramidnet_bottleneck' --dataset 'cifar100' --alpha 200 --depth 272 --num_classes 100 --epochs 300 --lr 0.5 --batch_size 128
Eval samples: 10000
图示cifar100训练验证过程
图5. 使用PyramidNet_bottleneck对cifar100的训练验证图示
使用PyramidNet网络进行动物分类的训练和验证
训练
In [16]
!python train.py --net 'pyramidnet'
In [15]
!python train.py --net 'pyramidnet_bottleneck'
验证
In [ ]
!python eval.py --net 'pyramidnet'
=> the layer configuration for each stage is set to [3, 3, 3, 3]
W0213 21:36:28.090073 12789 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0213 21:36:28.094696 12789 device_context.cc:372] device: 0, cuDNN Version: 7.6.
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 103/103 [==============================] - loss: 0.8536 - acc: 0.7357 - 187ms/step
Eval samples: 3276
{'loss': [0.8535936], 'acc': 0.7356532356532357}In [ ]
!python eval.py --net 'pyramidnet_bottleneck'
=> the layer configuration for each stage is set to [2, 2, 2, 2]
W0214 09:30:31.742739 30668 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0214 09:30:31.747535 30668 device_context.cc:372] device: 0, cuDNN Version: 7.6.
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 103/103 [==============================] - loss: 1.2630 - acc: 0.6734 - 194ms/step
Eval samples: 3276
{'loss': [1.2630014], 'acc': 0.6733821733821734}图示训练验证过程
图6. 使用PyramidNet的训练验证图示
使用resnet50网络进行动物分类的训练并验证
训练
In [14]
!python train.py --net 'resnet'
验证
In [ ]
!python eval.py --net 'resnet'
W0213 21:34:50.038996 12684 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0213 21:34:50.043457 12684 device_context.cc:372] device: 0, cuDNN Version: 7.6.
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 103/103 [==============================] - loss: 1.4232 - acc: 0.5888 - 191ms/step
Eval samples: 3276
{'loss': [1.4232028], 'acc': 0.5888278388278388}图示训练验证过程
图7. 使用ResNet的训练验证图示
比较
使用该项目的参数配置，即PyramidNet-32（alpha=300 without bottleneck）和PyramidNet-32（alpha=300 with bottleneck）相较于
resnet50模型具有更高的准确率，却具有少的多的参数。其中，resnet50的模型大小有141.5M，而PyramidNet-32（alpha=300 without bottleneck）
的模型大小有72.6M，PyramidNet-32（alpha=300 with bottleneck）的模型大小仅有48.9M。
图8. 使用PyramidNet和ResNet的训练验证比较图示
