https://github.com/yanqiangmiffy/tofu-vs-potato
美食识别挑战(1):豆腐VS土豆
https://github.com/yanqiangmiffy/tofu-vs-potato
Last synced: 9 months ago
JSON representation
美食识别挑战(1):豆腐VS土豆
- Host: GitHub
- URL: https://github.com/yanqiangmiffy/tofu-vs-potato
- Owner: yanqiangmiffy
- Created: 2019-11-29T09:52:25.000Z (almost 6 years ago)
- Default Branch: master
- Last Pushed: 2020-11-17T15:35:03.000Z (about 5 years ago)
- Last Synced: 2025-01-05T22:27:52.015Z (11 months ago)
- Language: Python
- Size: 2.78 MB
- Stars: 0
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# tofu-vs-potato
美食识别挑战(1):豆腐VS土豆
## CV分类学习
### 尝试1
使用VGG16/ResNet50/InceptionV3预训练模型,输入做了一些数据增强,效果不好,过拟合
线上得分:75.4537
### 尝试2
使用简单的CNN网络,效果不佳
线上得分:52
大佬建议1:
```text
@致Great 2层CNN可能不够哦
两层最多也就提取一些曲线特征
```
大佬建议2:
```text
参照vgg16多搞几层卷积
加入数据增强防止过拟合
```
### 尝试3
[阿水baseline](https://github.com/datawhalechina/competition-baseline/tree/master/competition/AI%E7%A0%94%E4%B9%A0%E7%A4%BE-%E7%BE%8E%E9%A3%9F%E8%AF%86%E5%88%AB%E6%8C%91%E6%88%98%EF%BC%881%EF%BC%89%EF%BC%9A%E8%B1%86%E8%85%90VS%E5%9C%9F%E8%B1%86 )
线上得分:98.8539
### 尝试4
keras-resnet50-tta
测试阶段进行数据增强:Test Time Augmentation
## 其他问题
- 使用os.listdir()存在文件名是按照字典序排序的,需要与label对应的文件专为一致
https://www.cnblogs.com/jins-note/p/9550388.html
- 不同型号的GPU会影响训练效果,显卡质量越高,计算精度越高