D2F - Description to Face Synthesis

A project to explore the use of transfer learning in GANs to produce photo realistic images of human faces from its description

Project breakdown

• Data preprocessing(effecient way to load/use the entire Celeb-A data onto the main memory)
• Language Model
• Implementing the Giant MSG-GANs on the entire LFW dataset
• Transfer Learning
• Final Model
• Accuracy Measure

Data preprocessing(effecient way to load/use the entire LFW data onto the main memory)

Requirement of the Data preprocessing code

• Modular
• Should work on atleast 2 levels of folder structure
  • Ex. images/example.png or images/folder1/example.png
• Should output one batchsize of images per training iteration
• Must preprocess the output batch images
• Low overhead
• Images shouldn't be stored in memory before and after the processing of the batch
• Should be Jupyter Notebook/Collab compatible

Language Model

Requirements

• Load(from pickle file,json or txt/csv ) ,preprpocess and encode
• Must be able to process text using both word level and sentence level embedding
  • Word level embedding Algorithms -
    • Glove,
    • Word2Vec,
    • FastText,
    • Generating sentence embedding from word embedding using -
      • Mean of all word emb
      • Min/Max/Mean of all the word emb
  • Sentence level embedding Algorithms -
  • Infersent,
  • Google Universal Encoder,
  • Elmo