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