Skip to content

Adversarial Discriminative Domain Adaptation with MNIST 64x64 in Lasagne-Theano

License

Notifications You must be signed in to change notification settings

Xun-Yang/adda_mnist64

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Adversarial Discriminative Domain Adaptation

Given the MNIST 64x64 handwritten recognition dataset, we define the following two sets as "source" and "target":

Source MNIST domain Target MNIST domain

The goal is to maximize the accuracy of a classifier on the "target" set. However, labels are only available for the "source" set. We formulate the task as a domain adaptation problem in adversarial terms. Two networks, the classifier and the domain discriminator, compete to optimize opposite objectives.

In particular, the discriminator predicts whether a sample image belongs to the "source" or the "target" domains, accessing features from the input images through the classifier network only. At the same time, the classifier has two simultaneous objectives: a) recognizing digits from the "source" domain in a supervised fashion, and b) fooling the discriminator by maximizing its classification error.

Source MNIST domain

The idea behind adversarial domain adaptation is that the classifier will eventually learn to hide features that are useful to discriminate between domains. By doing so, it becomes robust against domain differences and improves its classification accuracy in the "target" set.

Results

Classification accuracy:

Source Set Target Set Performance Gap
Without ADDA 0.99 0.38 0.61
With ADDA 0.99 0.78 0.21

Using adversarial domain adaptation, the performance gap between the "source" and "target" datasets decreases by 3-folds.

Usage

  • Simply execute python train.py

Requisites

References

About

Adversarial Discriminative Domain Adaptation with MNIST 64x64 in Lasagne-Theano

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%