and the true image, to see the difference

• Generated image:

• True image:

• Generated image:
  

• True image:

See Thoughts.md

• Finish generating embeddings
• Reduce learning rate before overfitting
• Add decaying dropout to the training
• Is DRAW outdated based on GAN recent results... DRAW architecture were every iteration is condition on previous generated images + new caption?
• Better data augmentation
• Make GAN works, whether WGAN or classical ones
• Refactor the code
• Try to generate "HD" images, like in StackGAN (need to retrieve from MS-COCO for supervised learning)
• Used ideas from there: