Generative Cooperative Network

Jan 1, 0001
Code
GCN Network Output

An implementation of the paper entitled “Generative Cooperative Net for Image Generation and Data Augmentation” by Qiangeng Xu et al, as a part of the final project for the course Deep Learning at the spring semester of 2021, University of Tehran.

DataSet

In this project we’ve use two different datasets:

Goals

This projects goal is to build a neural network to

  • Generat images of faces (for KDEF dataset)
  • Generat images of hadwritten number (for QMNIST dataset)
  • Create a new augmentation tool: After training the GCN network and combining the identity features of two people with a ratio of 0.5, new images can be produced

Training

With respect to different goals of hour network we train it on two different datasets as mentioned above. The input of networ for these datasets is different.

Input of Network for KDEF dataset

  • a one-hot encoded vector for identity of the image (with length of 70)
  • a one-hot encoded vector for face expression of the image (with length of 4)
  • a one-hot encoded vector for transformation of the number (with length of 8)
  • an image with size of 28*28

Input of Network for KDEF dataset

  • a one-hot encoded vector for number of the image (with length of 10)
  • a one-hot encoded vector for color the number (with length of 3 for R, G and B)
  • a one-hot encoded vector for transformation of the image (with length of 8)
  • an image with size of 158*158

Model

Our proposed model has two modules:

  • Generator: generate image with an MSE loss
  • Classifier: classifies the generated image of generator. In this structure, the generator and the classifier have goals in the same direction and they are trying to increase the quality of the produced images by working together. The Architecture of model is shown below
    GCN architecture
    model

Results

Here we have the results for our metioned tasks. right column of each image is model output.

  • Generating images of faces (for KDEF dataset)

    faces output
    model output faces

  • Generating images of handwritten numbers (for QMNIST dataset)

    numbers output
    model output numbers

  • Create a new augmentation tool

    augmentations output
    model output augmentation

Team Members

Deep Learning past
Amir Mesbah
Amir Mesbah
Master student in Computer Engineering (with a major in AI)

I am a graduated Master’s student from University of Tehran.