keras EfficientNet介绍,在ImageNet任务上涨点明显 | keras efficientnet introduction
本文首发于个人博客https://kezunlin.me/post/88fbc049/,欢迎阅读最新内容!
keras efficientnet introduction
Guide
About EfficientNet Models
compared with resnet50, EfficientNet-B4 improves the top-1 accuracy from 76.3% of ResNet-50 to 82.6% (+6.3%), under similar FLOPS constraint.
Using Pretrained EfficientNet Checkpoints
Keras Models Performance
- The top-k errors were obtained using Keras Applications with the TensorFlow backend on the 2012 ILSVRC ImageNet validation set and may slightly differ from the original ones.
The input size used was 224x224 for all models except NASNetLarge (331x331), InceptionV3 (299x299), InceptionResNetV2 (299x299), Xception (299x299),
EfficientNet-B0 (224x224), EfficientNet-B1 (240x240), EfficientNet-B2 (260x260), EfficientNet-B3 (300x300), EfficientNet-B4 (380x380), EfficientNet-B5 (456x456), EfficientNet-B6 (528x528), and EfficientNet-B7 (600x600).
notice
- Top-1: single center crop, top-1 error
- Top-5: single center crop, top-5 error
- 10-5: ten crops (1 center + 4 corners and those mirrored ones), top-5 error
- Size: rounded the number of parameters when
include_top=True
- Stem: rounded the number of parameters when
include_top=False
Top-1 | Top-5 | 10-5 | Size | Stem | References | |
---|---|---|---|---|---|---|
VGG16 | 28.732 | 9.950 | 8.834 | 138.4M | 14.7M | [paper] [tf-models] |
VGG19 | 28.744 | 10.012 | 8.774 | 143.7M | 20.0M | [paper] [tf-models] |
ResNet50 | 25.072 | 7.940 | 6.828 | 25.6M | 23.6M | [paper] [tf-models] [torch] [caffe] |
ResNet101 | 23.580 | 7.214 | 6.092 | 44.7M | 42.7M | [paper] [tf-models] [torch] [caffe] |
ResNet152 | 23.396 | 6.882 | 5.908 | 60.4M | 58.4M | [paper] [tf-models] [torch] [caffe] |
ResNet50V2 | 24.040 | 6.966 | 5.896 | 25.6M | 23.6M | [paper] [tf-models] [torch] |
ResNet101V2 | 22.766 | 6.184 | 5.158 | 44.7M | 42.6M | [paper] [tf-models] [torch] |
ResNet152V2 | 21.968 | 5.838 | 4.900 | 60.4M | 58.3M | [paper] [tf-models] [torch] |
ResNeXt50 | 22.260 | 6.190 | 5.410 | 25.1M | 23.0M | [paper] [torch] |
ResNeXt101 | 21.270 | 5.706 | 4.842 | 44.3M | 42.3M | [paper] [torch] |
InceptionV3 | 22.102 | 6.280 | 5.038 | 23.9M | 21.8M | [paper] [tf-models] |
InceptionResNetV2 | 19.744 | 4.748 | 3.962 | 55.9M | 54.3M | [paper] [tf-models] |
Xception | 20.994 | 5.548 | 4.738 | 22.9M | 20.9M | [paper] |
MobileNet(alpha=0.25) | 48.418 | 24.208 | 21.196 | 0.5M | 0.2M | [paper] [tf-models] |
MobileNet(alpha=0.50) | 35.708 | 14.376 | 12.180 | 1.3M | 0.8M | [paper] [tf-models] |
MobileNet(alpha=0.75) | 31.588 | 11.758 | 9.878 | 2.6M | 1.8M | [paper] [tf-models] |
MobileNet(alpha=1.0) | 29.576 | 10.496 | 8.774 | 4.3M | 3.2M | [paper] [tf-models] |
MobileNetV2(alpha=0.35) | 39.914 | 17.568 | 15.422 | 1.7M | 0.4M | [paper] [tf-models] |
MobileNetV2(alpha=0.50) | 34.806 | 13.938 | 11.976 | 2.0M | 0.7M | [paper] [tf-models] |
MobileNetV2(alpha=0.75) | 30.468 | 10.824 | 9.188 | 2.7M | 1.4M | [paper] [tf-models] |
MobileNetV2(alpha=1.0) | 28.664 | 9.858 | 8.322 | 3.5M | 2.3M | [paper] [tf-models] |
MobileNetV2(alpha=1.3) | 25.320 | 7.878 | 6.728 | 5.4M | 3.8M | [paper] [tf-models] |
MobileNetV2(alpha=1.4) | 24.770 | 7.578 | 6.518 | 6.2M | 4.4M | [paper] [tf-models] |
DenseNet121 | 25.028 | 7.742 | 6.522 | 8.1M | 7.0M | [paper] [torch] |
DenseNet169 | 23.824 | 6.824 | 5.860 | 14.3M | 12.6M | [paper] [torch] |
DenseNet201 | 22.680 | 6.380 | 5.466 | 20.2M | 18.3M | [paper] [torch] |
NASNetLarge | 17.502 | 3.996 | 3.412 | 93.5M | 84.9M | [paper] [tf-models] |
NASNetMobile | 25.634 | 8.146 | 6.758 | 7.7M | 4.3M | [paper] [tf-models] |
EfficientNet-B0 | 22.810 | 6.508 | 5.858 | 5.3M | 4.0M | [paper] [tf-tpu] |
EfficientNet-B1 | 20.866 | 5.552 | 5.050 | 7.9M | 6.6M | [paper] [tf-tpu] |
EfficientNet-B2 | 19.820 | 5.054 | 4.538 | 9.2M | 7.8M | [paper] [tf-tpu] |
EfficientNet-B3 | 18.422 | 4.324 | 3.902 | 12.3M | 10.8M | [paper] [tf-tpu] |
EfficientNet-B4 | 17.040 | 3.740 | 3.344 | 19.5M | 17.7M | [paper] [tf-tpu] |
EfficientNet-B5 | 16.298 | 3.290 | 3.114 | 30.6M | 28.5M | [paper] [tf-tpu] |
EfficientNet-B6 | 15.918 | 3.102 | 2.916 | 43.3M | 41.0M | [paper] [tf-tpu] |
EfficientNet-B7 | 15.570 | 3.160 | 2.906 | 66.7M | 64.1M | [paper] [tf-tpu] |
Reference
- tf efficientnet
- efficientnet keras pre-trained weights
- Implementation of EfficientNet model. Keras and TensorFlow Keras.
History
- 20190912: created.
Copyright
- Post author: kezunlin
- Post link: https://kezunlin.me/post/88fbc049/
- Copyright Notice: All articles in this blog are licensed under CC BY-NC-SA 3.0 unless stating additionally.
keras EfficientNet介绍,在ImageNet任务上涨点明显 | keras efficientnet introduction的更多相关文章
- Keras(一)Sequential与Model模型、Keras基本结构功能
keras介绍与基本的模型保存 思维导图 1.keras网络结构 2.keras网络配置 3.keras预处理功能 模型的节点信息提取 config = model.get_config() 把mod ...
- 解决 ImportError: cannot import name 'initializations' from 'keras' (C:\Users\admin\AppData\Roaming\Python\Python37\site-packages\keras\__init__.py)
解决 ImportError: cannot import name 'initializations' from 'keras' : [原因剖析] 上述代码用的是 Keras version: '1 ...
- 【Keras篇】---Keras初始,两种模型构造方法,利用keras实现手写数字体识别
一.前述 Keras 适合快速体验 ,keras的设计是把大量内部运算都隐藏了,用户始终可以用theano或tensorflow的语句来写扩展功能并和keras结合使用. 二.安装 Pip insta ...
- 深度学习利器: TensorFlow系统架构及高性能程序设计
2015年11月9日谷歌开源了人工智能平台TensorFlow,同时成为2015年最受关注的开源项目之一.经历了从v0.1到v0.12的12个版本迭代后,谷歌于2017年2月15日发布了TensorF ...
- Keras中图像维度介绍
报错问题: ValueError: Negative dimension size caused by subtracting 5 from 1 for 'conv2d_1/convolution' ...
- 关于keras框架的介绍以及操作使用
Keras 是一个 Python 深度学习框架,可以方便地定义和训练几乎所有类型的深度学习模型.Keras 最开始是为研究人员开发的,其目的在于快速实验.我们可以进入网站主页 - Keras 中文文档 ...
- keras系列︱Application中五款已训练模型、VGG16框架(Sequential式、Model式)解读(二)
引自:http://blog.csdn.net/sinat_26917383/article/details/72859145 中文文档:http://keras-cn.readthedocs.io/ ...
- 深度学习应用系列(二) | 如何使用keras进行迁移学习,以训练和识别自己的图片集
本文的keras后台为tensorflow,介绍如何利用预编译的模型进行迁移学习,以训练和识别自己的图片集. 官网 https://keras.io/applications/ 已经介绍了各个基于Im ...
- 【tf.keras】tf.keras加载AlexNet预训练模型
目录 从 PyTorch 中导出模型参数 第 0 步:配置环境 第 1 步:安装 MMdnn 第 2 步:得到 PyTorch 保存完整结构和参数的模型(pth 文件) 第 3 步:导出 PyTorc ...
随机推荐
- 2019/12/13学习内容摘要(Linux磁盘管理①)
一,查看磁盘或目录容量 1.命令df 查看已挂载磁盘的总容量,使用容量,剩余容量等,可以不加任何参数,默认以KB为单位 选项[-i] 表示查看inodes的使用情况 [-h] 表示用合适的单位显示 ...
- 2017 经典的CVPR 关于ImageCaptioning论文
1. SCA-CNN: Spatial and Channel-wise Attention in Convolutional Networks for Image Captioning ...
- gitlab-CI作业-yml
stages: - build - deploy before_script: - echo "Restore NuGet Packages..." - echo "do ...
- 10 个提升效率的Linux小技巧
您是否曾经惊讶于看到某人在 UNIX 中非常快速地工作,触发命令并快速地执行操作?是的,我碰到过几次,并且我一直都在向那些超级巨星开发者学习.在本文中,我想分享一些 UNIX 命令实践,这些实践是我在 ...
- np.array()和np.dot()的区别
1.生成数组的方式不同 2.乘法计算方式不同 array生成数组,np.dot()表示矩阵乘积,(*)号或np.multiply()表示点乘 mat生成数组,(*)和np.dot()表示矩阵相乘,点乘 ...
- C++如何使用宏定义来简化代码性能测试 | cpp macro like function to implement a performance profiler
本文首发于个人博客https://kezunlin.me/post/65dc693d/,欢迎阅读最新内容! cpp macro like function to implement a perform ...
- aop分层模型——aop是元编程的一种
织入应用层—->待织入的附加功能 应用层 织入管理层—->使用元语和织入的功能进行编程 语言层 aop元语层---〉aop机制提供的高阶抽象概念. 解释层 aop引擎层—->维护的实 ...
- JavaScript空字符串判断
JavaScript空字符串判断 本文完整示例代码GIT仓: 测试用例完整代码:isNullOrEmpty jPublic GIT仓:jPublic 比较常见写法 if (str == 'undefi ...
- node http 模块 常用知识点记录
关于 node,总是断断续续的学一点,也只能在本地自己模拟实战,相信总会有实战的一天~~ http 作为服务端,开启服务,处理路由,响应等 http 作为客户端,发送请求 http.https.htt ...
- Javascript 的定时器 setInterval,setTimeout,clearInterval
今天开通博客.来1个.哇哈哈哈~~ 今天本来想复习BOM的看到定时器也算DOM一种 ?(是这样吗).分享一下 参考源于:八神吻你 http://www.cnblogs.com/lmfeng/archi ...