GitHub - WuXunjin/MLPrototype: Simple prototypes of common ML concepts

Using Domain Adaptation for Few Shot Generative Modeling

Effect of Number of Instances on DC-GAN

FashionMNIST

Class	10	100	1000
T-Shirt/Top
Trouser
Pullover

MNIST

Class	10	100	1000
'1'
'2'
'3'

notMNIST

Class	10	100	1000
'B'
'C'
'D'

Effect of number of Instances of Target Domain on Cycle GAN

MNIST

Source	Target	10	100	1000
'1'	'7'
'3'	'8'
'6'	'9'

FashionMNIST

Source	Target	10	100	1000
Sneaker	Ankle Boot
Sneaker	Bag

notMNIST

Source	Target	10	100	1000
Sneaker	Bag

Evaluation

Confusion Matrix

Train Dataset	Test Dataset	10	100	1000
MNIST GAN	MNIST Real
FashionMNIST GAN	FashionMNIST Real
notMNIST GAN	notMNIST Real
MNIST Real	MNIST Real
FashionMNIST Real	FashionMNIST Real
notMNIST Real	notMNIST Real

Visualization

Train Dataset	10	100	1000
MNIST
FashionMNIST
notMNIST

Class/Domain distance

Within-domain distance

Dataset	Ex1	Ex2	Ex3
MNIST
FashionMNIST
CIFAR

Cross-domain distance

Primary Domain	Helper Domain	Ex1	Ex2	Ex3
MNIST	SVHN-BW
MNIST	USPS

Usage

Generating examples in Regular and Few-Shot Scenarios ( Training )

Execute the conditional DC-GAN with python cDCGAN_train.py. This would create 4 directories within DCGAN folder
- DCGAN/animation containing gif of generated images after every epoch
- DCGAN/loss containing loss values for both generator and discriminator vs. epochs
- DCGAN/plots containing a png file with generated images after last epoch
- DCGAN/models containing model file with discriminator and generator weights in pytorch format
- Below example with MNIST. [0,1,2] are few shot classes

Animation

Dataset	5000/1000	1000/100
MNIST
SVHN

Loss

Dataset	5000/1000	1000/100
MNIST
SVHN

Generating examples in Regular and Few-Shot Scenarios ( Testing )

- TBC

Computing Domain Distance

Execute the Maximum Mean Discrepancy comparison with python MMD_WD.py. This would create 1 directory withing DCGAN folder
- DCGAN/mmdValues containing histograms of average MMD values vs. Classes for a single domain (dataset)
Execute the MMD comparison with python MMD_CD.py. This would create 1 directory within DCGAN folder
- DCGAN/mmdValues containing histograms of average MMD values vs. Classes for two different domains (datasets)

Evaluation

Classification of test data with training data being one of the three
- Real data instances from original dataset
- Real data + Data generated by DC-GAN
- Real data + Data generated by DC-GAN using MMD distance as a proxy for training
- Classification using SVF with RBF kernel
Execute Evaluate/classify.py
- Primary class is a single class within a domain say MNIST
- Helper class is another fixed class within the same domain
- Results in Evaluate/MMD

MNIST	F-MNIST	CIFAR	SVHN

Execute Evaluate/classify_CD.py
- Primary class is a single class within a domain say MNIST
- Helper class is another class in a different helper domain say SVHN
- Results in Evaulate/crossDomainMMD

MNIST	F-MNIST	CIFAR	SVHN

Execute Evaluate/classify_all.py
- Primary class is a single class within a domain say MNIST
- Helper class are all the classes except the primary class withing same domain
- Results in Evaulate/MMDall

MNIST	F-MNIST	CIFAR	SVHN

Execute Evaluate/classify_all_CD.py
- Primary class is a single class withing a domain say MNIST
- Helper classes are all the classes in a different helper domain say SVHN
- Results in Evaulate/crossDomainMMDall

MNIST	F-MNIST	CIFAR	SVHN

References

MNIST
FashionMNIST dataset
notMNIST dataset
GAN Tutorial
CycleGAN Implementation
Siamese/Triplet Loss Implementation

TODO

Add classification results for DCGAN with/without MMD
DCGAN with MMD [ Learning from all classes of same dataset ]
MMD Comparison [ Cross and within domain ]
Batches with max and min MMD
Update .py files [presently in .ipynb format]
Sample generation in differnet directory from classificaiton

Name		Name	Last commit message	Last commit date
Latest commit History 100 Commits
CycleGAN		CycleGAN
DCGAN		DCGAN
Datasets		Datasets
Evaluate		Evaluate
MMD		MMD
NewExperiments		NewExperiments
Visualize/embedding		Visualize/embedding
data		data
siamese-triplet		siamese-triplet
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
myFirstGAN.py		myFirstGAN.py

License

WuXunjin/MLPrototype

Folders and files

Latest commit

History

Repository files navigation

Using Domain Adaptation for Few Shot Generative Modeling

Effect of Number of Instances on DC-GAN

FashionMNIST

MNIST

notMNIST

Effect of number of Instances of Target Domain on Cycle GAN

MNIST

FashionMNIST

notMNIST

Evaluation

Confusion Matrix

Visualization

Class/Domain distance

Within-domain distance

Cross-domain distance

Usage

Generating examples in Regular and Few-Shot Scenarios ( Training )

Animation

Loss

Generating examples in Regular and Few-Shot Scenarios ( Testing )

Computing Domain Distance

Evaluation

References

TODO

About

Resources

License

Stars

Watchers

Forks

Languages