FDEN

A COMPACT AND LEGIBLE SOURCE CODE IS COMING!
Implementation of Plug-in Factorization for Latent Representation Disentanglement.

Requirements

Tensorflow (2.0.0-alpha0)
Pillow

Data sets

1. Omniglot
   Download omniglot/python/images_background.zip and omniglot/python/images_evaluation.zip.
   Place them into data/raw/omniglot/ folder.
2. Mini-ImageNet
   Download ImageNet and rename it to images.zip.
   Download index files.
   Place them into data/raw/imagenet/ folder.
3. Oxford Flower 102
   Download images and labels.
   Place them into data/raw/oxford/ folder.
4. MS-Celeb-1M Low-shot
   NOTE: Official website is down. Please ask the organizers (Microsoft) for access to the data set.
   Download TrainData_lowshot.tsv, DevelopmentSet.tsv, TrainData_Base.tsv.
   Place them into data/raw/msceleb/ folder.

How to run

python main.py --dataset=0 --mode=100 --way=5 --shot=1
Dataset: 0 - Omniglot, 1 - MS-Celeb-1M, 2 - Mini-ImageNet, 3 - Oxford Flower
Mode:
Invertible Network: 0## Began, 1## ALI
Classifier: #0# Siamese Network, #1# Prototypical Network
Mode: ##0 Train Invertible network, ##1 Train FDEN
Example: 101 ALI invertible network + Siamese + train FDEN
Way, shot: C-way K-shot learning setting for training FDEN

settings.py

root_path: Root path
project_path: Source code path
raw_data_path: Raw data set path (~36GB)
preprocessed_data_path: Path to store preprocessed data (~110GB)
temp_data_path: Path to store temporary data (~115GB)
result_path = Path to save results (tensorboard event files, model weights, source code)