Text to Image Synthesizer using Generative Adversarial Networks

The purpose of this project is to train a generative adversarial network (GAN) to generate images from textual description of the image. For this particular project, I have used flower images from the Oxford 102 Flower Dataset.

Data:

The data of the image description was obtained from here. The image caption data link was obtained from the following github.

The flowers dataset has 102 categories of flower images. Each category has 40-258 images. The total number of flower images-description pairs used in this project is 8100.

The image descriptions were processed into character level embeddings using the 300 D GloVe embeddings. You can obtain them from here. I used the embeddings created on Wikipedia 2015 and Gigaword-5 data.

The images of the flowers are as below:

The captions are as below:

Architecture:

This is an experimental implementation of synthesizing images. The images are synthesized using the GAN-CLS Algorithm from the paper Generative Adversarial Text-to-Image Synthesis. However, we have not used Skip-Thoughts vectors, instead, we tried the implementation using the GloVe embeddings.

Image Source : Generative Adversarial Text-to-Image Synthesis Paper.

Note: We have tried to implement the GAN architecture as best as we could. It is not guaranteed to be the exact same as the one in the paper.

Implementation:

The iPython notebook has comments indicating what each cell does. You can then execute the cells in text_to_image_synthesizer_glove.ipynb to see how everything works.

The notebook has 5 parts:

• Data Pre-processing:
  • Converting the images into numpy arrays based on the pre-set pixel size and storing them.
  • Converting the image description into embeddings and storing them.
  • Sotring the captions in a csv file.
• Loading and Combining:
  • Loading all the images' numpy arrays and appending it to form the image data.
  • Loading the image description embedding numpy array.
• Data modeling:
  • Creating the discrimintor and generator networks.
  • Functions for calculating the discriminator and generator loss.
  • Setting the optimizer and learning rate.
• Training:
  • Function for the train step, that creates the images, calculates the loss and adjusts the gradients.
  • Function for the train, it fetches the batch data and passes it to train_step and collects all the metrics.
• Results:
  • Function for testing the output from the generator given an input noise and a caption

Result:

Note: Addtional references in the iPython notebook.