Before Getting Started

• Face transfer showed very limited success compared to the original paper. Please see our final report for a discussion of achievable results with this implementation.
• To generate the dataset of emoji images:
  • Create a directory called emoji_data in the current datasets directory
  • There are 2 emoji generation scripts: create_emojis.py is for creating small datasets (<1000 images) and create_emojis_parallel.py for large datasets.
  • Run python3 create_emojis.py or python3 create_emojis_parallel.py to generate the dataset. Note you must use python3 to run these scripts because of the url request.
  • To change the number of emojis generated, change the num_emojis variable at the top of the generation script.

CS 229 Final Project: Unsupervised Cross-Domain Image Generation

Davis Rempe, Xinru Hua, Haotian Zhang (Team 925)

Based on this paper.

Overview of files in this repo:

• /datasets/ - all dataset download/creation/processing scripts
• /pretrained_model/ - all saved pretrained models for f blocks
• Network for digit domain transfer:
  • digits_model.py
• Networks for face domain transfer:
  • faces_model.py - network as in paper
  • faces_model_v2.py - modified network for testing
  • net_sphere.py - SphereFace network (taken mostly from here)
  • open_face_model.py and SpatialCrossMapLRN_model.py - OpenFace network (taken mostly from here)
• Training scripts:
  • base_test.py - abstract training class
  • classifier_f_test.py - f block training for digit transfer
  • digit_model_test.py - digit model transfer training
  • digit_model_test_septrain.py - digit model transfer training with separated source and target training
  • faces_model_test_open.py - face transfer training with OpenFace
  • faces_model_test_sphere.py - face transfer training with SphereFace
• Training script drivers:
  • FaceMain.ipynb - driver for training face transfer
• data.py - all data loaders and preprocessing code

Other references

• digit model architecture inspired by this implementation
• training strategies inspired by this implementation