Reconstruction of Generative Adversarial Networks in Cross Modal Image Generation with Canonical Polyadic Decomposition

Algorithm 1

Overall scheme of compact architecture training algorithm.

Input: mini-batch images , matching text , mismatching text , number of training batch steps

Output: a compact architecture for text-to-image GAN

1: Obtain three small layers as using Equation (5),(6),(7) to decompose original convolutional layer;

2: Adopt autoencoder to pre-train model layer by layer;

3: Select an appropriate learning rate for the decomposed model;

4: for    to  do

5: Encode matching text description and mismatching text description to description embedding ;

6: Draw sample of random noise ;

7: Concatenate to description embedding ;

8: Feed forward through generator and generate samples of {real image, right text}, {real image, wrong text} and {fake image, right text};

9: Update discriminator D using Adam;

10: Update generator G using Adam;

11: end for