Chris Pollett > Students >
Lei

    ( Print View)

    [Bio]

    [Blog]

    [CS 297 Proposal]

    [GANs-PDF]

    [Gen Videos Scene Dynamics-PDF]

    [3D CNNs Human Action RECOG-PDF]

    [TGAN-PDF]

    [Deliverable 1]

    [Deliverable 2]

    [Deliverable 3]

    [Deliverable 4]

    [CS297 Report - PDF]

    [CS 298 Proposal]

    [CS 298 Report - PDF]

    [CS 298 Slides - PDF]

Implement a simple GAN with Python and use it to generate Chinese character digits

Description: 

I choose this project to get familiar with GAN. After training a self-collected Chinese characters numbers from 0 to 9, we expect to get a batch of fake numbers in the same style. I use deep convolutional GAN and self-collected training dataset.


Dataset:

Numbers in Chinese character : 0 to 9

1) In the beginning, I used two sets of hand-writing numbers, which seems too hard to get a better result because of thin strokes

2) Then I chose to use thick strokes numbers

Use the augmentation.py to expand the training dataset to 10,000 pictures.

Download train.zip



Deep Convolutional GAN:

Discriminator - With three convolutional layers and one fully-connected layer.

Generator - With three convolutional layers and one fully-connected layer.

Use Adam optimizer and fixed learning rate 0.0002 and momentum term of 0.8.

Use batch normalization in both discriminator and generator and leaky ReLU.

Download dcgan.py


Output:

Here is the result after training 20000 epochs.

result



Conclusion:

A high-quality training dataset is critical for GAN.