Skip to content

Implementation of Deep Embedded Validation for Domain Adpatation on visda2017 dataset with MCD model

Notifications You must be signed in to change notification settings

Chen-Dixi/MCD_with_DEV

Repository files navigation

MCD_with_DEV

Implementation of Deep Embedded Validation for Domain Adpatation on visda2017 dataset with MCD model

Required

  • python3.6
  • pytorch 1.2.0
  • dixitool
  • tqdm
  • numpy

Intro

在VisDA2017数据集,Maximum Classifier Discrepancy模型上实现DEV模型选择方法,

DEV中的density ratio通过训练一个输出probability scalar的域判别器来计算得到

只是在深度DomainAdaptation模型上 对DEV进行粗略的实现,计算论文中的无偏估计

  • 训练出MCD模型,在main.sh里调整超参,把数据集的路径修改为自己的
bash main.sh
  • 使用数据集中的validation split和test split训练域判别器,判别器的模型在MCD/models.py中,参考SinGAN中的域判别器,用卷积对3x224x224的图片进行下采样。大概采用这样的方法,但是这个域判别器的效果一般😅
bash train_discriminator.sh
  • 用前两步训练的模型,在数据集的validation split上面计算相对于test split的无偏估计
bash DEV.sh

About

Implementation of Deep Embedded Validation for Domain Adpatation on visda2017 dataset with MCD model

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published