eternal-bugshine

This repository houses methods related to feature visualization of models trained on beetles.

Structure of the repository

• All core functionalities are stored in the src folder.
  • src/utils.py stores modules implementing utilities for handling datasets, training models and visualizing results.
    • src/utils/datasets.py implements functionality for reading datasets, getting dataset statistics, splitting datasets and performing data augmentations.
    • src/utils/headers.py contains definitions of constants, in particular the imagenet dataset statistics and default parameters for the deepdream method
    • src/utils/training.py contains all functionality related to model training
    • src/utils/visual.py contains functions for visualizing tensors and plotting multiple system on the same graph
  • src/deep_dream_aux.py implements functionality used by the deepdream algorithm such as functions for calculating scale-space levels, smoothing gradients and converting to and from tensors
  • src/deep_dream.py implements functions for implementing deep dream scale space and gradient ascent
  • src/models.py contains classes exposing layers in different models, currently ResNet50 and GoogleNet.
• For a tutorial covering all functionality check tutorial.ipynb.
• To play with the deep dream algorithm use playground.ipynb.
• Note that the repo is a work in progress! This means there are most likely still many bugs present.
  • currently you can only run the code from the two Jupyter notebooks.
  • GoogleNet does not support training from scratch