Learning Rotation Invariant Features for Cryogenic Electron Microscopy Images

Learning to seperate pose and staracture using GAN.

Setup

This is python3.6 and Pytorch based code. Dependencies:

# Using conda. If package installation fails it install with pip.
conda install --yes pip
conda config --add channels anaconda
conda config --add channels conda-forge
conda config --add channels pytorch
while read requirement; do conda install --yes $requirement || pip install $requirement; done < requirements.txt 
 
# Using pip
pip install -r requirements.txt 

Datasets

Datasets as tarballs are available from the links below.

• 5HDB simulated EM images

Download and extract. Working directory stracture:

.
├── LICENSE
├── README.md
├── configs
│   └── vae_mnist.yaml
├── data
│   └── 5HDB
│   └── MNIST
├── externals
│   └── spatial_vae
├── models
├── output
├── requirements.txt
└── src

Usage

Training spatial-VAE model:

cd src
python main_train_vae.py --config_path=../configs/vae_mnist.yaml
python main_train_vae.py --config_path=../configs/vae_5hdb.yaml

configuration file is located here: ''configs/vae_mnist.yaml"

Training our approach

cd src
python main_train_ours.py --config_path=../configs/ours_mnist.yaml 
python main_train_ours.py --config_path=../configs/ours_5hdb.yaml

configuration file is located here: ''configs/ours_mnist.yaml"

Pretrained models:

https://drive.google.com/file/d/1HoFbyV8I8AwBNFtlURWzrwdzk1__Asmx/view?usp=sharing

Ablation experiments

The most basic method

python main_train_ours.py --config_path=../configs/ours_5hdb.yaml architecture=fc use_wasserstein=false 

Using DCGAN

python main_train_ours.py --config_path=../configs/ours_5hdb.yaml architecture=cnn use_wasserstein=false 

DCGAN + Wasserstein

python main_train_ours.py --config_path=../configs/ours_5hdb.yaml architecture=cnn use_wasserstein=true 

License

This source code is provided under the MIT License.