This project reimplements SV2P network and its depended sub-networks in PyTorch-1.0.0.

Denote the project home directory as $PROJ_HOME.

First of all, create an virtual environment named rt:

cd "$PROJ_HOME"
python3 -m virtualenv rt
. rt/bin/activate
pip install -r requirements.txt

Next, if you have direnv installed, cd to $PROJ_HOME should automatically brings out prompt to direnv allow the .envrc file, which prepares necessary environment variables. Otherwise, do . "$PROJ_HOME/.envrc".

Under $PROJ_HOME/experiments/cdna there's a main.py as a launcher for CDNA-Net training. By python main.py --help, it accepts an ini configuration file and an action. For example,

[dataset]
dataset_name=MovingMNIST
indices=(range(8000), range(8000, 9950), range(9950, 10000))
in_channels=1
cond_channels=0

[train]
n_masks=2
batch_size=16
lr=0.001
max_epoch=10
seqlen=20
criterion_name=DSSIM

[train_device]
device=cuda

The key names are one-to-one corresponding to the positional/keyword argument names of sv2p.cdna.CDNA.__init__. It should be fairly straightforward what each key means.

An example run:

python main.py my-run.ini train

will produces my-run.ini.log as log file and runs-${TODAY_DATETIME}/ as checkpoint/statistics/visualization base directory.

To use dataset:

• MovingMNIST: download MovingMNIST and put it under $PROJ_HOME/data/MovingMNIST/ (make directory if not exists)