Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

口红? #4

Open
zhaipro opened this issue Apr 22, 2020 · 18 comments
Open

口红? #4

zhaipro opened this issue Apr 22, 2020 · 18 comments

Comments

@zhaipro
Copy link
Owner

zhaipro commented Apr 22, 2020

为什么又开始玩口红了?

@zhaipro
Copy link
Owner Author

zhaipro commented Apr 22, 2020

s

@zhaipro
Copy link
Owner Author

zhaipro commented May 8, 2020

版本 loss dice_coeff psnr val_loss val_dice_coeff val_psnr
2.0.0 0.0047 0.8735 29.3643 0.0046 0.7376 28.0056
2.0.1 0.0045 0.8829 29.6047 0.0042 0.6776 27.3260
2.0.2 0.0868 0.9132 28.2183 0.1161 0.5191 24.912
2.0.3 0.0894 0.9184 28.9648 0.0572 0.7362 27.1730
2.0.4 0.0868 0.9203 30.5728 0.1535 0.5305 24.6504
2.0.5 0.1170 0.8935 28.7663 0.0746 0.4990 25.1946
2.0.6 0.1085 0.8997 31.2613 0.0854 0.6298 26.0387
2.0.7 0.0942 0.9128 31.9750 0.0581 0.9050 29.6066
2.0.7.1 0.1038 0.9051 31.2448 0.0573 0.9143 29.9382
2.0.7.3 0.1027 0.9063 31.3688 0.0578 0.9173 29.9247
2.0.7.4 0.1044 0.9049 31.2056 0.0257 0.9250 29.5189
2.0.7.5 0.0930 0.9180 52.0262 0.0662 0.9085 29.2761
2.0.8 0.0929 0.9139 32.0576 0.0607 0.8951 29.5029
2.0.8.1 0.0990 0.9118 0.0443 0.9270
2.0.8.2 0.0901 0.9201 0.0747 0.9252
2.0.8.3 0.0928 0.9176 0.0714 0.9275
2.0.8.4 0.0917 0.9220 0.0702 0.9275
2.0.8.5 0.0879 0.9252 0.0696 0.9303
2.0.9 0.0929 0.9140 32.0216 0.0581 0.8966 29.6190
2.1.1 0.1201 0.8895 30.2393 0.0753 0.8790

在colab上表现良好,可为什么在我本机却过拟合?

@zhaipro
Copy link
Owner Author

zhaipro commented May 8, 2020

看来损失函数很重要呀

@zhaipro
Copy link
Owner Author

zhaipro commented May 13, 2020

洗牌后的效果:
loss: 0.0016 - dice_coeff: 0.7766 - psnr: 35.3927 - val_loss: 6.2160e-04 - val_dice_coeff: 0.6873 - val_psnr: 35.8081

@zhaipro
Copy link
Owner Author

zhaipro commented May 19, 2020

  • 嘴唇:[80, 50, 230]
  • 脸:[190, 90, 90]
  • 舌头:[0, 255, 255]

@zhaipro
Copy link
Owner Author

zhaipro commented May 21, 2020

惨痛的经验:开发环境还是交给conda吧,我不行啦,对不起pip

@zhaipro
Copy link
Owner Author

zhaipro commented May 23, 2020

数据增强已经足够了,只要不会过拟合就行,现在的目标是正确率啦。

@zhaipro
Copy link
Owner Author

zhaipro commented May 23, 2020

代办项

  1. 损失函数
  2. 批量大小
  3. 模型深度
  4. 模型广度
  5. Dropout

@zhaipro
Copy link
Owner Author

zhaipro commented May 26, 2020

数据强化

  • 水平翻转
  • 缩放
  • 挪动
  • 模糊
  • 色彩抖动
  • 侧脸
  • 遮挡,蒙眼
  • 老人
  • 大眼睛?
  • 根据复杂度打分

只要数据强,其他的都是小事儿。

@zhaipro
Copy link
Owner Author

zhaipro commented May 27, 2020

Dropout能否帮助模型寻找最优解?

@zhaipro
Copy link
Owner Author

zhaipro commented May 28, 2020

代办项

  • 评测XX的分割力度

没办法,我们的分割力度能到90%就不错啦,不求100分,但求最高分。

@zhaipro
Copy link
Owner Author

zhaipro commented May 28, 2020

根据人脸扭转的程度来表示训练权重?

@zhaipro
Copy link
Owner Author

zhaipro commented May 29, 2020

最令我开心的事儿就是发现自己是个脑残…

@zhaipro
Copy link
Owner Author

zhaipro commented Jun 2, 2020

起始层不要提取太多的信息,不然容易过拟合。

@zhaipro
Copy link
Owner Author

zhaipro commented Jun 4, 2020

3970

@zhaipro
Copy link
Owner Author

zhaipro commented Jun 4, 2020

hair cloth lip
96164 / 85977983 = 0.0011319825261650242

@zhaipro
Copy link
Owner Author

zhaipro commented Jun 5, 2020

720

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant